Human–AI Collaboration and Intellectual Provenance Record

1. Purpose and Scope

This record documents the development of Why Trust Oreskes’s Plumber? and reconstructs, to the extent supported by the surviving evidence, the respective roles of Daniel Capitanu and OpenAI’s ChatGPT in the conception, research, organization, structural and argumentative development, drafting, revision, and editorial preparation of the work.

Its purpose is evidentiary. It seeks to determine how the project originated and developed; what intellectual, structural, analytical, textual, and editorial contributions were made during that development; when relevant contributions or decisions occurred; and, where the evidence permits, how their provenance can reasonably be attributed.

The inquiry concerns the production process of the work as a whole rather than merely the immediate origin of particular sentences. The developmental record indicates that the book emerged through an extended human–AI process involving intellectual conception, research and preparatory documentation, chapter and section architecture, drafting instructions, AI-assisted analytical and textual development, human evaluation, rejection and redirection, substantive revision, integration, version control, and final editorial authorization. These stages did not necessarily occur identically in every chapter or passage, and this record does not assume that they did.

The record therefore examines several dimensions separately where the evidence permits: intellectual origin; research and source selection; pre-drafting organization; conceptual and argumentative development; structural design; textual generation; substantive revision; selection and arrangement; and final editorial control. The fact that one participant generated the wording of a passage does not, by itself, determine the provenance of every concept, argument, constraint, source choice, structural decision, or later transformation embodied in that passage. Conversely, the origin of a concept or the exercise of editorial control does not, by itself, establish the provenance of the resulting expression.

A central purpose of the record is consequently to preserve distinctions that simplified descriptions of human–AI production may obscure. The documented process is not adequately represented either as a completed human manuscript subsequently polished by artificial intelligence or as an autonomously generated AI manuscript produced in response to a general request. The evidence instead shows substantial contributions on both sides of the developmental relationship, differing in kind, timing, function, and degree across the project.

This record forms the third level of a four-level transparency architecture.

Level 1 — Notice: the published book contains a concise disclosure of material AI involvement.

Level 2 — Explanation: the publisher’s public website provides an accessible account of the nature of the human–AI collaboration.

Level 3 — Accountability: the present record provides the methodological reasoning, developmental reconstruction, representative evidence, and provenance analysis supporting those public descriptions.

Level 4 — Verification: a separate restricted evidentiary archive preserves underlying primary materials capable of supporting more detailed examination where a legitimate verification need exists.

Level 3 is therefore intended to be public and sufficiently detailed to permit meaningful evaluation of the reconstruction without requiring publication of the complete private developmental archive. Its function is not merely to disclose that artificial intelligence was used, but to explain with evidentiary discipline how it was used, how human and AI contributions interacted, and what the surviving developmental history supports concerning the intellectual provenance of the resulting work.

The record is deliberately proportionate rather than exhaustive. Its purpose is to establish a sufficiently supported account of the production process, not to reproduce every conversation, draft, intermediate file, or editorial decision generated during the development of an 823-page work. More granular primary documentation remains relevant because it permits the reconstruction to be tested where greater specificity is reasonably required.

The governing objective is therefore not maximal attribution.

It is maximal accuracy.

2. What This Record Does — and Does Not — Establish

This record reconstructs intellectual provenance and the documented production process of Why Trust Oreskes’s Plumber?. It identifies evidence concerning human intellectual initiative, continuing human direction, substantive human intellectual contribution, substantial AI participation, textual generation, revision, selection, and final editorial control. It also identifies contributions that arose through iterative interaction and, where the evidence does not permit reliable attribution, leaves the matter unresolved rather than supplying a retrospective attribution.

The record does not assign a percentage of authorship to Daniel Capitanu or ChatGPT. The developmental process cannot meaningfully be represented by a division such as a percentage of words, ideas, or labor. Different dimensions of provenance may follow different paths: a concept may originate with one participant, be substantially developed by the other, receive an AI-generated formulation, undergo human redirection or revision, and enter the published work through a final act of selection and arrangement. A single numerical allocation would erase rather than clarify those distinctions.

The record does not presume that the participant who produced the final wording of a sentence necessarily originated the intellectual material embodied in it. Nor does it presume that the participant who originated a concept necessarily produced or legally authored every later expression through which that concept was developed.

The record does not treat prompting, conversation length, repeated interaction, final approval, or selection among outputs as automatic proxies for substantive human authorship. Those acts may be relevant, but their evidentiary significance depends upon the intellectual decisions they embody and their effect upon the resulting work.

Similarly, the record does not treat every substantial activity performed by ChatGPT as equivalent to authorship. Some documented functions resemble research assistance, analytical assistance, developmental editing, continuity management, fact-checking, bibliographic support, or other conventional forms of literary and scholarly assistance. Other contributions involved the development of arguments, structures, counterarguments, analogies, formulations, and substantial potentially publishable prose. The record distinguishes among these functions rather than collapsing them into either “mere assistance” or “co-authorship.”

The record does not designate ChatGPT as a legal author or co-author.

It does not establish, by declaration, the legal ownership of the work.

It does not determine the copyright status of every sentence, paragraph, section, or other expressive element initially generated through artificial intelligence.

It does not convert human intellectual direction, selection, revision, or final editorial control automatically into human authorship of expression that may have originated elsewhere.

It likewise does not treat the immediate generation of prose by ChatGPT as dispositive of the provenance of the underlying intellectual project, research selection, conceptual framework, argumentative constraints, structural organization, or subsequent transformation of that prose.

The expression AI-enhanced authorship, used later in this record, is a descriptive term developed for purposes of characterizing the documented production process. It is not presented as an established category of Canadian copyright law, as an official publishing classification, or as a substitute for any definition of AI-generated or AI-assisted content imposed by a publisher, distributor, platform, registration system, legislature, court, or administrative authority.

Accordingly, describing the work globally as an instance of AI-enhanced authorship does not imply that individual passages cannot also qualify as AI-generated under an applicable external definition. The descriptions operate at different analytical levels.

The classification also does not imply that AI participation necessarily improved the literary, scholarly, factual, or intellectual quality of the book. In this context, “enhanced” refers functionally to the expansion of the author’s working capacities through substantial AI participation in research, analysis, organization, argumentative development, drafting, critical testing, rewriting, and revision. It is not an evaluative claim that the resulting work became better merely because artificial intelligence participated in its production.

Nor does this record claim exhaustive forensic attribution of every component of the work. The evidentiary basis consists instead of longitudinal developmental evidence, contemporaneous records, preparatory documentation, manuscript history, repeated evidence of human and AI intervention, canonical production records, and representative granular reconstruction of significant portions of the development process.

The distinction is important:

sufficient evidence to characterize the documented production process

is not the same undertaking as

exhaustive attribution of every intellectual and expressive component of the work.

The former is the purpose of this record. The latter is neither claimed nor required for the descriptive conclusions reached here.

Finally, the record does not ask the reader to accept the author’s retrospective account merely because it is the author’s account. Contemporary records receive priority wherever they survive. Later recollection may identify questions for investigation but is not silently transformed into contemporaneous evidence. Likewise, evidence favorable to human intellectual provenance is not discounted merely because it benefits the author, and evidence establishing substantive AI contribution is not minimized merely because it complicates a simpler account of human authorship.

The standard is symmetrical.

Where the evidence establishes a human contribution, the record attributes it accordingly.

Where the evidence establishes an AI contribution, the record attributes it accordingly.

Where the contribution was developed iteratively and cannot reasonably be reduced to either participant acting independently, the record describes the developmental sequence.

Where the evidence is genuinely insufficient, the record says so.

Within these limits, the record establishes a documented account of the intellectual provenance and production process of Why Trust Oreskes’s Plumber?. Legal conclusions concerning copyright, ownership, or the status of particular expressive elements remain distinct questions to be determined, where necessary, under the applicable legal and institutional standards.

3. Method of Reconstruction

The reconstruction undertaken in this record is historical and evidentiary. Its purpose is to determine, as accurately as the surviving materials permit, how Why Trust Oreskes’s Plumber? developed and how different forms of human and artificial-intelligence contribution entered that development.

The governing methodological principle is:

Primary Evidence Before Retrospective Description.

Where contemporaneous developmental records survive, they receive priority over later recollection or retrospective characterization of the process. Later descriptions may identify questions requiring investigation, clarify the significance attributed to earlier events, or assist in locating relevant evidence. They are not silently treated as equivalent to records created during the development itself.

Relevant evidence includes, where available:

• dated human–AI conversations;
• preparatory research and planning documents;
• Continuity Briefs and Analytical Dossiers;
• chapter- and section-level development records;
• source-selection and source-verification records;
• author instructions, corrections, and constraints;
• proposed, rejected, redirected, and revised structures or formulations;
• successive manuscript drafts and intermediate versions;
• placement and integration records;
• canonical print, EPUB, and other production files;
• version-control records;
• editorial and production records capable of establishing chronology; and
• later records capable of corroborating earlier developmental events.

No category of evidence is treated as self-interpreting. A final manuscript, for example, is strong evidence of what ultimately appeared in the work and of how the work was represented at that stage. It is not, by itself, evidence of where the underlying concepts, arguments, structures, or expressions originated. Conversely, a conversational record showing that ChatGPT generated a passage establishes immediate textual generation but does not necessarily establish the provenance of every concept, source choice, argumentative constraint, structural decision, or later revision embodied in the resulting passage.

3.1 Unit of Analysis

The individual prompt is not treated as the presumptive unit of provenance analysis.

A substantial portion of the development of Why Trust Oreskes’s Plumber? involved intellectual and documentary work occurring before prose generation. Chapter-level and later section-level planning records were used to establish such matters as:

• the function of a chapter or section within the larger argument;
• questions requiring examination;
• relevant sources and evidence;
• conceptual distinctions;
• argumentative constraints;
• continuity requirements;
• relationships with preceding and subsequent chapters;
• anticipated footnote and bibliography requirements;
• internal architecture;
• page strategy;
• risks requiring particular attention; and
• conditions that had to be satisfied before drafting proceeded.

Accordingly, the relevant production sequence may extend well beyond:

human prompt → AI output.

Where supported by the surviving evidence, the more appropriate analytical sequence is:

intellectual conception → research and preparatory documentation → chapter architecture → section-level planning → drafting instructions → AI analytical and textual contribution → human evaluation → acceptance, rejection, redirection, or revision → integration → version control → final editorial authorization.

Not every chapter or section is presumed to have followed this sequence identically. The reconstruction determines from the evidence when particular stages existed, how they functioned, and who contributed to them.

3.2 Pre-Drafting Evidence

Contemporaneous records demonstrably predating the corresponding prose receive particular attention because they may establish the existence of intellectual decisions before textual generation occurred.

Such records may show that a research problem, conceptual distinction, source selection, argumentative boundary, chapter function, structural relationship, or editorial constraint already existed before ChatGPT was asked to produce the relevant prose.

This evidentiary significance must not be overstated.

The existence of a preparatory document does not establish that its contents originated exclusively with the human participant. Preparatory documentation may itself have been substantially organized, developed, or worded by ChatGPT.

Its own provenance must therefore be reconstructed.

The relevant question is not simply:

Who generated the planning document?

It is:

Where did the intellectual components organized within that document originate, how were they developed, and who exercised consequential judgment concerning their inclusion, modification, or rejection?

Where the surviving evidence permits, preparatory material is therefore distinguished as:

HUMAN — introduced or materially determined by Daniel Capitanu;

AI — introduced or materially developed by ChatGPT;

ITERATIVE — developed through interaction in a manner that cannot reasonably be reduced to either participant acting independently; or

UNRESOLVED — insufficiently documented for reliable attribution.

3.3 Multiple Dimensions of Provenance

The reconstruction does not assume that provenance is a single variable.

Where relevant, it examines separately:

intellectual origin — where the underlying problem, concept, thesis, or conceptual framework arose;

research and source selection — who identified, proposed, selected, or required particular evidence or sources;

pre-drafting organization — how existing materials were organized before prose generation;

conceptual development — who introduced or materially transformed concepts and distinctions;

argumentative development — how arguments, counterarguments, qualifications, and implications were developed;

structural design — how the book, chapters, sections, and argumentative sequence were organized;

textual generation — who or what produced particular formulations or passages;

substantive revision — how generated or existing material was materially changed;

selection and arrangement — which material was retained, rejected, relocated, combined, or reordered; and

final editorial control — who determined the form ultimately authorized for publication.

These dimensions may have different provenance within the same passage.

A concept may originate with the human participant, receive substantial AI analytical development, be expressed initially in AI-generated prose, undergo later human revision, and enter the final manuscript through human selection and arrangement.

Where such a sequence is supported by the evidence, the record describes the sequence rather than forcing all dimensions into a single attribution.

3.4 Evidentiary Neutrality and Competing Explanations

The reconstruction adopts a position analogous to that of an independent fact-finder.

Its purpose is neither to establish human authorship at all costs nor to construct the strongest possible challenge to it.

The adversarial component of the analysis is a test of evidentiary strength, not a presumption against any participant.

For material provenance questions, competing explanations are considered where the evidence reasonably permits them.

The governing question is:

Which account of the developmental process is best supported by the surviving evidence considered as a whole?

The methodology therefore seeks to avoid two symmetrical errors.

Confirmation bias occurs when ambiguous evidence is interpreted preferentially in favor of a desired account of human intellectual provenance.

Disconfirmation bias occurs when ordinary evidence of human intellectual contribution is subjected to an artificially skeptical standard merely because artificial intelligence participated substantially in the project.

The same evidentiary standard applies in both directions.

Where contemporaneous evidence establishes that Capitanu introduced a concept, imposed an argumentative constraint, selected a research direction, rejected an alternative, redirected an architecture, or made another consequential intellectual decision, that contribution is attributed to him without artificial qualification.

Where contemporaneous evidence establishes that ChatGPT introduced a concept, developed an argument, proposed an architecture, generated a formulation, or produced expressive material subsequently retained in the work, that contribution is attributed to ChatGPT without minimization.

Where the evidence supports a genuinely iterative development, the record describes it as such rather than imposing a false binary attribution.

Where multiple reconstructions remain reasonably possible, the competing explanations may be stated and their relative evidentiary support assessed.

3.5 Standard of Attribution

This record does not require proof beyond reasonable doubt.

Nor does it assume that every intellectual act occurring during a months-long creative process must possess an independently preserved timestamped document before it can be recognized.

Contemporaneous documentation receives the greatest evidentiary weight. Attribution may also be supported by:

• multiple independent or temporally distinct records;
• manuscript history;
• pre-drafting documents;
• repeated developmental patterns;
• version relationships;
• later records that reliably corroborate earlier events; and
• other evidence reasonably capable of supporting the reconstruction.

Retrospective recollection receives less weight than contemporary evidence but is not automatically disregarded. It may identify matters requiring investigation or contribute to a reconstruction where appropriately characterized and corroborated.

The absence of a surviving document does not prove that an event did not occur.

Likewise, the existence of a later statement does not transform that statement into contemporaneous evidence of the event it describes.

The applicable standard is therefore neither maximum skepticism nor maximum attribution.

It is the most accurate reconstruction reasonably supported by the surviving record.

3.6 Symmetrical Counterfactual Analysis

Where useful, the reconstruction employs a functional comparison with conventional human literary, research, editorial, or drafting assistance.

The question is not simply:

Could this AI contribution be characterized as authorship?

Nor is it:

Could this contribution be characterized as mere assistance?

The symmetrical question is:

If a human collaborator had performed the same function under materially comparable conditions of intellectual direction, independence, expressive contribution, revision, and decision authority, how would that contribution most reasonably be characterized?

The answer may differ by function.

Research organization may resemble research assistance.

Source verification may resemble fact-checking.

Structural diagnosis may resemble developmental editing.

Conversion of author-defined intellectual requirements into a detailed working plan may resemble literary or analytical assistance.

Independent introduction of a consequential argument may resemble substantive intellectual collaboration.

Generation of substantial publishable prose creates a distinct expressive-provenance question.

This test is an analytical control rather than a legal rule. Its purpose is to avoid both exaggerating AI contribution merely because it was performed by an AI system and minimizing substantial expressive or intellectual contributions merely because the human participant retained final authority.

3.7 Proportionality of Reconstruction

The evidentiary method is designed to characterize the production process reliably without requiring sentence-by-sentence reconstruction of the entire work.

The record therefore combines longitudinal evidence across the project with selected granular reconstructions capable of testing whether the broader characterization is supported.

Representative analysis is used to examine:

• the state of the intellectual project before generation;
• the provenance of preparatory documentation;
• substantial AI developments;
• human interventions during production;
• rejection and redirection;
• textual and structural revision;
• and the relationship between developmental records and the canonical manuscript.

Where a representative reconstruction reveals a pattern inconsistent with the general account, that inconsistency must be incorporated rather than suppressed.

The method therefore distinguishes between:

sufficient evidence to characterize the production process

and

exhaustive forensic attribution of every component of the work.

The former is the purpose of the present record.

The latter is not presumed necessary to establish a responsible account of the project’s intellectual provenance.

4. Evidentiary Vocabulary

The following vocabulary is used throughout this record to distinguish the strength of evidence, the kind of contribution being examined, and the degree to which provenance can reasonably be attributed.

These terms are analytical classifications used within this dossier. They do not themselves determine legal authorship or copyright status.

4.1 Evidentiary Status

DOCUMENTED

A finding is DOCUMENTED when it is directly supported by an identifiable contemporaneous record capable of establishing the material fact for which it is cited.

Examples may include:

• a dated conversation in which a concept is introduced;
• a pre-drafting document establishing a constraint before prose generation;
• a manuscript version containing the relevant material at a known stage;
• a documented rejection or redirection;
• or a production record establishing the authorized state of the manuscript.

DOCUMENTED does not mean that every surrounding question has been resolved.

A document may establish that an event occurred while leaving the provenance of subordinate elements unresolved.

CORROBORATED

A finding is CORROBORATED when more than one independent, temporally distinct, or otherwise meaningfully separate element of the surviving record supports the same reconstruction.

Corroboration may occur, for example, where:

• a concept appears in a developmental conversation and later in preparatory documentation;
• an earlier constraint is reflected consistently in later manuscript versions;
• multiple records independently show the persistence of a structural decision;
• or later evidence confirms a developmental event without being used as proof of its original provenance.

CORROBORATED is not a weaker synonym for DOCUMENTED. A finding may be directly documented in one respect and corroborated in another.

INFERRED

A finding is INFERRED when it is reasonably supported by the surviving evidence but is not directly established by a complete contemporaneous record.

Inference may be appropriate where:

• documentary sequences strongly support a developmental relationship but an intermediate record is missing;
• surviving versions allow a reasonable conclusion concerning change over time;
• or several evidentiary elements collectively support a reconstruction without directly recording the relevant intellectual act.

An inference must be identified as such.

It is not upgraded to DOCUMENTED merely because it is plausible or consistent with later recollection.

UNRESOLVED

A matter is UNRESOLVED when the surviving evidence is insufficient to determine provenance, responsibility, chronology, or developmental relationship with reasonable confidence.

UNRESOLVED is not a presumption against human contribution, against AI contribution, or against collaborative development.

Nor is it a designation to be used merely because absolute certainty is unavailable.

It is reserved for genuine evidentiary insufficiency.

Where the evidence supports one account more strongly than competing accounts, the appropriate evidentiary classification should reflect that support rather than default automatically to UNRESOLVED.

4.2 Provenance Attribution

The following terms identify the presently best-supported source or developmental character of a contribution.

HUMAN

A contribution is classified as HUMAN when the evidence supports attribution of the relevant intellectual, structural, editorial, or expressive contribution to Daniel Capitanu.

Depending upon the issue, this may include:

• initiating a problem;
• introducing a concept or distinction;
• selecting or requiring a research direction;
• imposing an argumentative constraint;
• rejecting a proposed conclusion;
• determining a structural relationship;
• introducing a governing principle;
• substantially rewriting material;
• or making another consequential intellectual or editorial decision.

HUMAN does not imply that ChatGPT made no later contribution to the same material.

AI

A contribution is classified as AI when the evidence supports attribution of the relevant contribution to ChatGPT.

This may include:

• introducing a development not found in the prior human input;
• proposing a structural architecture;
• expanding an argument;
• generating a counterargument;
• identifying an implication;
• producing a formulation or analogy;
• generating substantial prose;
• reorganizing material;
• or performing another consequential analytical, developmental, or expressive function.

AI attribution does not imply that the resulting material entered the published work unchanged or without human direction, evaluation, revision, or authorization.

ITERATIVE

A contribution is classified as ITERATIVE when the surviving evidence shows that it developed through human–AI interaction and cannot reasonably be reduced to either participant acting independently.

This may occur where:

• a human concept receives substantial AI development;
• an AI proposal is materially redirected by the human participant and subsequently redeveloped;
• repeated cycles of proposal, criticism, revision, rejection, and reintegration transform the material;
• or the resulting concept, architecture, or formulation is the product of successive contributions whose separation would distort the developmental history.

ITERATIVE does not mean that the individual components of the sequence are unknowable.

Where possible, the record may still distinguish, for example:

governing concept — HUMAN

analytical development — AI

subsequent restructuring — HUMAN

final formulation — ITERATIVE or AI

The classification describes the relevant unit at the level being examined.

UNRESOLVED PROVENANCE

A contribution has UNRESOLVED PROVENANCE where the record does not support reliable attribution to HUMAN, AI, or ITERATIVE development.

This classification should be used only where the evidentiary deficiency is material to the provenance question.

4.3 Functional Vocabulary

The following terms distinguish kinds of contribution that may coexist within the same developmental sequence.

CONCEPTION

The formation or initiation of the intellectual project, problem, question, thesis, conceptual framework, or other originating intellectual element.

Conception is distinct from later development and from textual expression.

RESEARCH

The identification, location, examination, comparison, organization, or verification of sources, evidence, historical material, scientific literature, or other documentary inputs relevant to the work.

Research provenance may itself be divided between source proposal, source selection, source interpretation, and ultimate use.

DEVELOPMENT

The substantive expansion, refinement, testing, qualification, or transformation of an existing intellectual element.

Development may concern concepts, arguments, counterarguments, implications, distinctions, analogies, or other intellectual material.

STRUCTURAL DEVELOPMENT

The organization of material at the level of the book, part, chapter, section, argumentative sequence, transition, or other compositional architecture.

Structural provenance is analytically distinct from the provenance of the ideas organized within that structure.

DRAFTING

The production of prose intended or potentially suitable for incorporation into the manuscript.

Drafting identifies expressive production. It does not, by itself, establish the origin of the concepts or arguments expressed.

REDRAFTING

The production of a new or substantially reworked textual version of previously existing material.

Redrafting may involve both expressive and intellectual change and must therefore be examined according to the nature of the revision.

REVISION

A substantive alteration of existing material.

Revision may include correction, qualification, expansion, compression, restructuring, replacement, or other changes affecting meaning, argument, organization, or expression.

Routine typographical correction is not treated as equivalent to substantive revision.

SELECTION

The decision to retain one element rather than another.

Selection may be editorially significant but is not treated automatically as authorship of the selected expression.

Its evidentiary importance depends upon the intellectual judgment embodied in the selection.

REJECTION

The deliberate exclusion of a proposed concept, argument, structure, formulation, passage, or other developmental element.

Rejection is particularly relevant where it reveals substantive boundaries governing the work.

REDIRECTION

An intervention that materially changes the objective, conclusion, argumentative path, structure, treatment, or intended function of developing material.

Redirection differs from simple approval or correction because it alters the course of development.

INTEGRATION

The incorporation of developed material into the wider architecture and continuity of the manuscript.

Integration may require decisions concerning placement, relationship to surrounding material, transitions, redundancy, consistency, and argumentative function.

EDITORIAL CONTROL

Authority exercised over the state ultimately authorized for publication, including decisions to accept, reject, redirect, revise, reorganize, remove, or retain material.

Editorial control is relevant to the production history of the work.

It is not treated as automatically equivalent to authorship of every expressive component controlled.

INTELLECTUAL PROVENANCE

The developmental history of the intellectual components of the work, including their origin, transformation, interaction, organization, expression, revision, selection, and final disposition.

Intellectual provenance is broader than immediate textual generation and distinct from a legal determination of copyright ownership.

TEXTUAL PROVENANCE

The developmental origin and transformation of particular expressive material.

Textual provenance may include initial generation, subsequent rewriting, combination with other material, modification, and final incorporation.

It is analytically distinct from the provenance of the underlying idea or argument.

FINAL AUTHORIZATION

The human decision that a particular manuscript state or production version is authorized to represent the work.

Final authorization establishes publication control.

It does not retrospectively change the provenance of material incorporated into the authorized version.

4.4 Relationship Between Evidentiary and Provenance Classifications

Evidentiary status and provenance attribution answer different questions.

DOCUMENTED / CORROBORATED / INFERRED / UNRESOLVED describe the strength or character of the evidentiary basis.

HUMAN / AI / ITERATIVE / UNRESOLVED PROVENANCE describe the presently best-supported developmental attribution.

The two classifications may therefore be combined.

Examples include:

DOCUMENTED — HUMAN

A contemporaneous record directly establishes a human contribution.

DOCUMENTED — AI

A contemporaneous record directly establishes an AI-originating contribution.

CORROBORATED — ITERATIVE

Multiple records support the conclusion that a development emerged through interaction.

INFERRED — HUMAN

The available evidence reasonably supports human provenance, but the originating event is not directly preserved.

UNRESOLVED

The evidence is insufficient to support reliable attribution.

This two-axis vocabulary prevents evidentiary confidence from being confused with the identity of the contributor.

4.5 Governing Rule

The classifications in this record are tools for disciplined reconstruction, not rhetorical labels.

They must not be upgraded to strengthen a preferred conclusion or downgraded merely because a well-supported attribution creates difficulty for another part of the analysis.

Where evidence establishes a contribution, the record should say so.

Where evidence supports an iterative sequence, the sequence should be described.

Where the evidence supports only an inference, it should remain an inference.

Where the record cannot reliably determine provenance, the matter should remain unresolved.

The vocabulary exists to preserve those distinctions consistently throughout the remainder of the dossier.

5. Production Model

The development of Why Trust Oreskes’s Plumber? did not follow a simple recurring sequence in which the author supplied a prompt and ChatGPT returned prose for acceptance or rejection.

The surviving record supports a substantially more structured production model.

By late May 2026, a recurring pre-drafting protocol had been established in which relevant working materials were consulted and the intellectual and editorial requirements of the next unit were defined before substantial prose generation proceeded.

The production process developed around a set of standing and chapter-specific working materials that included, at different stages:

• a Continuity Brief, used to preserve the governing framework, accumulated decisions, terminology, structural relationships, and continuity across a large manuscript;
• Analytical Dossiers, used to organize research questions, conceptual distinctions, source requirements, argumentative problems, risks, and page objectives;
• Naomi Oreskes’s Why Trust Science? and other primary or secondary source materials relevant to the unit under development;
• the Manuscript Draft So Far, used to monitor continuity, duplication, integration, established terminology, and the relationship between new and existing material;
• chapter- and section-specific planning records;
• source and bibliography requirements;
• footnote claims;
• page and structural targets;
• placement and transition instructions;
• and later version-control and integration records.

These materials were not merely logistical aids.

They formed part of the intellectual production environment in which individual chapters and sections were developed.

5.1 Separation of Planning and Drafting

A significant feature of the production model was the explicit separation between preparatory analysis and prose generation.

Before drafting, the working process could require identification of:

• the function of the relevant chapter or section within the wider book;
• its governing question or thesis;
• the internal argumentative sequence;
• relevant analytical dossiers;
• sources requiring consultation or verification;
• distinctions that had to be preserved;
• conclusions that could or could not legitimately be drawn;
• continuity with prior chapters;
• the intended relationship to later chapters;
• anticipated footnote requirements;
• preliminary bibliography;
• page allocation;
• sensitive argumentative or evidentiary issues;
• and risks of distortion, repetition, premature conclusion, or imbalance.

Drafting was not automatically authorized merely because a topic had been identified.

The surviving record documents a working discipline in which the preparatory stage was expected to be reviewed, validated, or redirected before substantial prose generation proceeded.

This distinction is important because it separates two activities that can otherwise be collapsed into the generic word “prompting”:

defining the intellectual task

and

executing the textual development of that task.

5.2 Recurrence Across the Manuscript

The pre-drafting system was not confined to a single chapter.

The surviving developmental record identifies successive preparatory checkpoints associated with major portions of the book, including:

• Part II, Chapters 3–4;
• Part III, Chapters 5–6;
• Part IV, Chapters 7–8;
• Part V, Chapters 9–10;
• Part VI, Chapters 11–12;
• Part VII, Chapters 13–15; and
• Part VIII, Chapters 16–17.

The precise contents and provenance of individual planning records differ and must not be generalized beyond the evidence.

The recurrence of the system is nevertheless significant.

Across successive stages of the project, the process repeatedly incorporated an identifiable phase of research organization, continuity control, conceptual and structural planning, and validation preceding or accompanying substantial drafting.

The production model is therefore better represented as:

standing project documentation → chapter-specific analysis → human–AI development of a pre-drafting architecture → human validation, correction, or redirection → authorization to draft → substantial AI and iterative development → prose generation → human evaluation → further research, rejection, revision, restructuring, or redirection → integration → version control → final editorial authorization.

This model describes a recurring production pattern rather than a rigid mechanical sequence.

Not every chapter followed every stage in the same manner or to the same degree.

5.3 Provenance Within the Production Model

The production system itself was collaborative.

Its existence must not be mistaken for proof that all preparatory architecture was independently created by Daniel Capitanu.

The record shows substantial ChatGPT participation in preparing planning memoranda, organizing analytical material, proposing structures, developing argument sequences, identifying implications, and transforming accumulated inputs into executable drafting plans.

At the same time, surviving records identify consequential human contributions within the pre-drafting process.

These include, among others:

• requiring continued use of the documentary and continuity system;
• requiring validation before drafting;
• imposing distinctions that the architecture had to preserve;
• defining argumentative boundaries;
• limiting when controversial material could enter the book;
• determining how opposing positions were to be treated;
• rejecting or replacing proposed architectures;
• defining relationships among chapters;
• and determining the role that particular parts of the book were to play in the overall argument.

The relevant provenance question is therefore not:

Who wrote the planning memorandum?

It is:

Who supplied, introduced, developed, selected, constrained, reorganized, validated, or rejected the intellectual components embodied in the planning process?

A planning document generated substantially by ChatGPT may still preserve documented human intellectual inputs.

Conversely, the fact that the author approved such a document does not convert AI-originating analytical or structural contributions into human-originating contributions.

5.4 Human Intervention Before Generation

The developmental record contains examples showing that pre-drafting human participation was substantive rather than merely procedural.

For the development of Chapters 3–4, Capitanu required the project to preserve a distinction between methodological and ontological naturalism and to avoid turning the early discussion prematurely into a confrontation over Intelligent Design.

For later discussion of conflicts of interest, he required examination not only of financial interests but also of prestige, reputation, career incentives, and institutional authority.

During development of the later architecture of the book, an initial analytical arrangement was not simply accepted. Capitanu imposed a different governing structure in which Chapters 13–17 were to operate as a five-movement dramatic sequence, Chapter 15 as the climax, and the final part as a philosophical and civic resolution.

These interventions occurred upstream of subsequent prose generation.

They therefore provide evidence of intellectual direction and architectural decision-making that cannot be reduced to retrospective selection among completed AI outputs.

5.5 AI Participation Within the Production Model

ChatGPT's role within this system was likewise substantive.

It participated in:

• organizing research;
• identifying source needs;
• producing and maintaining analytical working records;
• proposing chapter and section architecture;
• converting author-defined problems and constraints into developed plans;
• identifying argumentative implications;
• generating counterarguments;
• testing claims adversarially;
• proposing transitions and reorganizations;
• producing substantial prose;
• redrafting;
• identifying inconsistencies or redundancies;
• assisting with source verification and bibliographic work;
• and supporting revision and production control.

The production model therefore cannot accurately be described either as mechanical AI execution of a complete human plan or as autonomous AI construction of the work from general instructions.

It was a structured iterative system in which different kinds of intellectual, analytical, expressive, and editorial contribution entered at different stages.

5.6 Validation, Redirection, and Authorization

Human validation played a recurring role in the production sequence.

Validation did not imply that every element of a plan or draft was human-originating.

Its function was to determine whether development could proceed within the governing objectives and constraints of the project.

Depending upon the stage, the author could:

• authorize drafting;
• require additional research;
• reject a proposed structure;
• relocate an argument;
• narrow or broaden a treatment;
• impose a suspended conclusion;
• require symmetry or fairness;
• forbid a premature inference;
• demand further qualification;
• require revision;
• or determine that a resulting unit was acceptable for integration.

Generated material was therefore not treated as self-authorizing merely because it existed.

The distinction between generation and authorization remained operative throughout the project and later continued into canonical version control and format-specific production.

5.7 Production Model Finding

The surviving evidence does not support reducing the development of Why Trust Oreskes’s Plumber? to:

human prompt → AI-generated book.

Nor does it support:

human-written book → AI editorial polishing.

The best-supported production model is instead one of sustained iterative development involving:

human-originating intellectual objectives and constraints; collaborative research and pre-drafting organization; substantial AI analytical, structural, and expressive contribution; continuing human evaluation and redirection; iterative revision and integration; and final human editorial control over the published form.

This production model provides the procedural context within which the more specific provenance questions examined in the following sections must be understood.

6. Intellectual Origin

The question of intellectual origin is distinct from the later questions of architecture, drafting, argument development, and textual expression.

For purposes of this record, the relevant inquiry is not whether every important argument in the final book originated with Daniel Capitanu.

The evidence does not support so broad a claim.

The more precise question is whether the project itself, its initiating intellectual problem, its central critical orientation, and its core intellectual framework originated in identifiable human intellectual activity before substantial AI development transformed that framework into the resulting book.

The surviving record supports that conclusion.

6.1 Formation of the Project

The earliest presently identified development records date from 26–27 May 2026.

By that stage, the project was already centered upon Naomi Oreskes’s treatment of trust in science and upon the plumber analogy associated with that discussion.

The emerging project did not begin by rejecting Oreskes’s argument for informed trust.

Its central move was more specific.

The case for informed trust would be taken seriously, but the mechanisms proposed to warrant that trust would themselves be treated as objects of inspection.

The underlying question was therefore not simply:

Should science be trusted?

It was closer to:

What makes trust in science warranted, and may the mechanisms invoked to justify that trust themselves be inspected?

This orientation became foundational to the book.

6.2 Initial Critical Field

The early human-defined inquiry already extended beyond a single objection or controversy.

The recovered record identifies an emerging field of examination that included:

• philosophical assumptions;
• scientific consensus;
• peer review;
• conflicts of interest;
• historical failures;
• selective appeals to technological success;
• institutional authority;
• scientific expertise;
• and the conditions under which public trust is warranted.

This did not yet constitute the final book architecture.

It did, however, establish the intellectual terrain from which that architecture would subsequently be developed.

6.3 Rhetorical and Epistemic Orientation

Capitanu also established an early rhetorical constraint that remained important throughout the project.

The work was to operate as a restrained or sober satire, not as an anti-scientific polemic.

This was not merely a stylistic preference.

It had epistemic consequences.

The project was to criticize mechanisms, assumptions, institutional practices, and boundaries without treating criticism of science as equivalent to rejection of science.

Oreskes’s argument was not to be caricatured simply to create an easier target.

Later development repeatedly preserved this requirement that opposing positions receive fair treatment even when subjected to critical inspection.

6.4 The Plumber Motif

The plumber motif was already present at the earliest recovered stage of the project and became the central organizing metaphor of the book.

Its intellectual function was not simply comic.

The plumber analogy supplied a way to translate an abstract problem of trust and expertise into a practical question of inspectability:

a citizen may reasonably rely upon expertise while still asking what warrants that reliance and whether the relevant procedures, judgments, and institutional mechanisms remain open to examination.

The precise first appearance of every component of the eventual title and every intermediate formulation cannot be established from all presently preserved primary records.

The evidentiary claim should therefore remain appropriately limited.

What the surviving record supports is that the plumber framework was already integral to the project during its earliest documented formation and became the durable vehicle through which the broader trust problem was developed.

6.5 Intellectual Origin Versus Book Architecture

A major provenance distinction appears almost immediately.

On 27 May, ChatGPT proposed an organized seventeen-chapter architecture that transformed the developing field of questions into a book-length progression.

That proposal constituted a substantial AI contribution.

It organized material involving Oreskes, consensus, philosophy, historical examples, peer review, technological achievement, contested inference, contemporary scientific problems, and public trust into a coherent structural sequence.

Capitanu accepted the seventeen-chapter structure as the working architecture and subsequently directed its modification and development.

This event is central to the provenance reconstruction because it distinguishes two things that could easily be conflated:

intellectual initiation and critical orientation of the project

from

initial book-level architecture.

The presently recovered record supports the first as human-originating and the second as involving substantial AI contribution.

The fact that ChatGPT made a consequential architectural contribution does not erase the prior intellectual problem from which the architecture was constructed.

Conversely, the human origin of the intellectual problem does not justify retrospectively attributing the seventeen-chapter architecture entirely to the human participant.

6.6 Early Argumentative Constraints

The early drafting stage provides additional evidence that the project possessed human intellectual direction before substantial prose generation.

Capitanu restricted the initial drafting phase to the Prologue and Chapters 1–2 and directed that Intelligent Design and anti-Intelligent-Design material not be introduced prematurely into that portion of the work.

He also required fair treatment of Oreskes rather than construction of her position as a rhetorical target.

ChatGPT subsequently produced substantial prose within those constraints.

This sequence demonstrates an early separation between:

human intellectual and argumentative direction

and

AI textual execution and development.

It does not imply that every later concept within those chapters was human-originating.

It does establish that the earliest drafting occurred within an already defined intellectual and rhetorical jurisdiction.

6.7 Early Conceptual Inputs

Other early records further support identifiable human intellectual contribution.

One example concerns Jim Baggott’s Farewell to Reality and the expression “The Authorized Version.”

Capitanu introduced Baggott into the developing discussion and connected the notion of an authorized scientific account with questions concerning consensus and epistemic diversity.

The resulting distinction concerned the difference between:

an institutionally authorized account of reality

and

truth itself.

ChatGPT subsequently participated in developing and integrating this distinction into the broader argument.

This example is representative of a recurring pattern:

human-introduced intellectual input → substantial AI development → iterative integration into the manuscript.

6.8 Development Beyond the Originating Framework

The finding of human intellectual origin must not be overstated.

The final book contains important developments to which ChatGPT contributed substantially at the level of:

• architecture;
• analysis;
• argument;
• counterargument;
• conceptual development;
• analogy;
• formulation;
• and prose generation.

Some developments were proposed initially by ChatGPT and subsequently retained, revised, relocated, or integrated by Capitanu.

The evidence therefore does not support the proposition that all substantive arguments in the finished work originated with the author.

For this reason, the terminology adopted in this record is deliberately:

core intellectual framework

rather than

core arguments.

The former accurately describes the evidence.

The latter would collapse later AI-originating or AI-developed arguments into the initial human provenance of the project.

6.9 Intellectual-Origin Finding

The surviving developmental record supports the following distinction:

The conception, intellectual direction, and core intellectual framework of Why Trust Oreskes’s Plumber? originated with Daniel Capitanu and were developed under his continuing direction.

At the same time:

ChatGPT made substantial subsequent contributions to the architecture, research, structural and argumentative development, drafting, critical testing, rewriting, and revision through which that intellectual project became the resulting book.

These propositions are not contradictory.

They concern different dimensions and stages of provenance.

The intellectual-origin finding therefore does not depend upon claiming that Capitanu independently devised every argument, structure, formulation, or passage.

It rests upon the documented emergence of a human-originating intellectual problem and critical framework, followed by a sustained developmental process in which artificial intelligence contributed substantially to converting that framework into the published work.

7. From Concept to Architecture

The transition from the originating intellectual problem of Why Trust Oreskes’s Plumber? to the architecture of a full-length book was one of the earliest points at which substantial AI contribution became visible.

The surviving developmental record distinguishes between:

the intellectual initiation and critical orientation of the project

and

the initial organization of that project into a seventeen-chapter book structure.

The former is supported as human-originating.

The latter involved a substantial contribution from ChatGPT.

7.1 The Initial Seventeen-Chapter Architecture

On 27 May 2026, during the early development of the project, ChatGPT proposed an organized seventeen-chapter architecture.

The proposal transformed a developing field of intellectual questions into a sustained book-length progression involving, among other subjects:

• Naomi Oreskes and the problem of scientific trust;
• consensus;
• philosophical assumptions;
• historical cases;
• peer review;
• technological success;
• institutional and methodological boundaries;
• contested inference;
• contemporary problems in scientific practice;
• and public trust.

The importance of this contribution should not be minimized.

The seventeen-chapter proposal did more than arrange headings.

It converted a set of intellectual problems into an extended argumentative sequence capable of supporting a book-length treatment.

This is therefore classified as a substantial AI architectural contribution.

Capitanu accepted the seventeen-chapter structure as the working architecture and subsequently directed its development, modification, and internal reorganization.

The resulting book should therefore not be described as possessing an exclusively human-originating architecture.

Nor does the evidence support the opposite description in which ChatGPT independently conceived both the intellectual project and the structure through which it was expressed.

The documentary record instead supports:

human intellectual origin and critical framework → substantial AI book-level architecture → continuing human direction and iterative architectural development.

7.2 Architecture as a Developmental Framework, Not a Fixed Outline

The initial seventeen-chapter structure did not function as a static outline that was simply filled with prose.

As the project developed, individual chapters acquired more precise functions, relationships, boundaries, and internal architectures.

New distinctions entered the project.

Some arguments were relocated.

Some proposed placements were rejected.

Some chapter functions were narrowed or expanded.

Some developments required restructuring of later parts of the book.

The original seventeen-chapter architecture therefore provided continuity at the level of the book while remaining subject to substantial internal development.

This distinction is important.

Structural continuity does not imply structural immobility.

The final manuscript retained the seventeen-chapter form, but the intellectual function of individual chapters and the relationships among them evolved significantly during production.

7.3 Human Constraints on Early Architecture

Human architectural direction appears almost immediately after adoption of the initial structure.

Capitanu restricted the earliest drafting phase to the Prologue and Chapters 1–2.

He further required that Intelligent Design and anti-Intelligent-Design sources not be introduced prematurely into this part of the book.

This was not merely a request concerning subject matter.

It established an architectural rule concerning argumentative sequence.

Certain controversies were to be delayed until the intellectual groundwork necessary to examine them had been established.

Oreskes’s position was likewise to receive fair treatment rather than be introduced merely as a rhetorical target.

The architecture therefore developed under constraints concerning:

• sequencing;
• admissibility of material at particular stages;
• fairness to opposing positions;
• avoidance of premature conclusions;
• and preservation of the larger argumentative trajectory.

7.4 Architecture Through Conceptual Differentiation

As the manuscript developed, architectural decisions increasingly reflected conceptual distinctions that had to be preserved across chapters.

One important example concerns methodological and ontological naturalism.

The developing book required these to remain analytically distinct rather than allowing methodological rules governing scientific practice to be treated automatically as ontological conclusions about what kinds of causes exist.

This distinction affected not only argument content but chapter placement and sequencing.

Likewise, the treatment of Intelligent Design developed as a problem of admissibility and methodological inspection before becoming a problem of substantive evidentiary evaluation.

The architecture therefore came to embody a division of intellectual labor:

one chapter could examine the conditions under which an inference is admitted or excluded;

another could examine the substantive strength of the inference itself.

The distinction between procedural admissibility and epistemic victory thus became architectural as well as conceptual.

7.5 Architecture as Iterative Decision-Making

The surviving record contains several instances in which architectural proposals moved through distinct stages of contribution.

A representative pattern is:

human intellectual question → AI development and proposed placement → human relocation or redirection → iterative integration.

One example appears in the development of the “philosophical charity” and “inspectable trust” material.

Capitanu raised the question whether science requires the kind of philosophical charity often requested for it or whether warranted trust requires inspection.

ChatGPT developed this intervention into a proposed section and further developed the language of inspectable trust.

ChatGPT initially proposed placing the material in Chapter 17.

Capitanu redirected it to Chapter 15 in order to preserve Chapter 17 as the concluding synthesis.

The sequence is therefore:

initiating question — HUMAN

development and formulation — AI / ITERATIVE

initial placement — AI

final architectural placement — HUMAN

This example illustrates why architectural provenance cannot be reduced to the identity of the participant who first proposed a section heading or placement.

7.6 Development of the Later Book Architecture

The architecture became increasingly deliberate in the later parts of the manuscript.

By July, the relationship among Chapters 13–17 had become a distinct architectural problem.

An analytical plan for Chapters 16–17 was not simply accepted.

Capitanu superseded it with a broader governing architecture in which Chapters 13–17 were to function as a five-movement dramatic unit.

Within that arrangement:

• the earlier chapters prepared the accumulated institutional and epistemic tensions;
• Chapter 15 was assigned the role of climax;
• and Part VIII, containing Chapters 16–17, was to function as philosophical and civic resolution.

This intervention demonstrates that architectural control operated above the level of isolated chapters.

The author was not merely accepting or rejecting local AI-generated structures.

He was determining the dramatic and argumentative function of multi-chapter sequences within the work as a whole.

ChatGPT nevertheless continued to contribute substantially to the internal organization and development of those units.

7.7 Continuity of the Seventeen-Chapter Form

By mid-to-late August, production records show a mature seventeen-chapter structure.

The basic numerical architecture proposed in May had survived.

Individual chapters, however, had undergone substantial development in:

• internal organization;
• argumentative function;
• conceptual content;
• placement of examples;
• relationship to neighboring chapters;
• and role within the larger trajectory of the book.

The evidence therefore establishes continuity of the seventeen-chapter framework without implying continuity of every subordinate structural element.

The surviving production structure is best understood as the result of:

an AI-proposed initial macro-architecture

followed by

months of iterative human–AI structural development under continuing human editorial and intellectual direction.

7.8 Architectural Provenance Finding

The evidence does not support either of two simplified descriptions.

It does not support:

Daniel Capitanu independently designed the complete book architecture and ChatGPT merely filled it with prose.

The initial seventeen-chapter organization was a substantial ChatGPT contribution.

Nor does it support:

ChatGPT independently designed the book while Capitanu merely approved the resulting structure.

The architecture developed within a human-originating intellectual project and was repeatedly constrained, redirected, reorganized, and functionally redefined through human intervention.

The most accurate reconstruction is:

human-originating intellectual framework → substantial AI macro-architectural proposal → human acceptance and governing constraints → iterative chapter- and section-level development → repeated human redirection and structural decision-making → mature seventeen-chapter canonical architecture.

This architectural history forms the bridge between the originating intellectual framework described in the preceding section and the research and pre-drafting system through which individual chapters were subsequently developed.

8. Research and Pre-Drafting Development

Research and pre-drafting development constituted a distinct and recurring stage in the production of Why Trust Oreskes’s Plumber?.

The surviving record does not support treating the manuscript as a sequence in which topics were supplied directly to ChatGPT and converted immediately into finished prose.

Instead, successive parts of the book were preceded by preparatory work involving research organization, continuity control, chapter function, structural planning, source requirements, argumentative constraints, risks, footnote needs, bibliography, page strategy, and validation before drafting.

The importance of this stage is both developmental and evidentiary.

Developmentally, it organized the materials from which prose would later be generated.

Evidentially, it permits examination of what intellectual structure existed before the corresponding prose appeared.

8.1 Establishment of the Pre-Drafting Protocol

By 27 May 2026, the project was already operating with an explicit expectation that drafting should be preceded by structured preparation.

Before substantial prose generation, the relevant unit could require clarification of:

• function;
• thesis or governing question;
• internal architecture;
• analytical dossiers;
• source requirements;
• footnote claims;
• preliminary bibliography;
• continuity considerations;
• page strategy;
• and sensitive argumentative issues.

Drafting was expressly separated from this preparatory stage.

The working discipline required validation before prose generation proceeded.

This is significant because it establishes that pre-drafting was not invented retrospectively as an explanation of how the project was supposedly developed.

It appears within the developmental record itself as an operative production method.

8.2 Standing Documentary Environment

As the project expanded, research and planning became organized around recurring working resources.

These included:

Continuity Brief

The Continuity Brief served to preserve accumulated project-level decisions and prevent later chapters from becoming detached from the conceptual and editorial framework already established.

Its functions included maintaining consistency in:

• terminology;
• prior conclusions;
• chapter relationships;
• argumentative boundaries;
• unresolved questions;
• and lessons emerging from earlier drafting.

By 3 June, an updated Continuity Brief was being used specifically to transfer lessons from the Prologue and Chapters 1–2 into later stages of development.

Analytical Dossiers

The Analytical Dossiers functioned as research and conceptual reservoirs.

Depending on the part of the book under development, they could contain or organize:

• research questions;
• relevant sources;
• distinctions requiring analysis;
• possible arguments and counterarguments;
• historical or scientific cases;
• evidentiary weaknesses;
• risks of overstatement;
• page targets;
• and material potentially relevant to multiple chapters.

Primary and Secondary Sources

The project made use of Oreskes’s source material and a wider research field extending into philosophy of science, scientific history, methodological naturalism, peer review, institutional authority, Intelligent Design, contemporary research problems, and other subjects.

The relevant point for provenance purposes is not simply that sources were consulted.

It is that source selection and research organization occurred as identifiable developmental activities before or alongside drafting.

Existing Manuscript

The manuscript developed cumulatively.

The existing draft therefore became itself a working source.

New material had to be tested against what already existed for:

• repetition;
• contradiction;
• continuity;
• placement;
• terminology;
• argumentative progression;
• and integration.

8.3 Documented Pre-Drafting Checkpoints

The surviving record identifies successive pre-drafting stages across most of the book.

Part II — Chapters 3–4

By 3 June 2026, preparation included:

• page targets;
• architecture;
• analytical constraints;
• dossier use;
• provenance considerations;
• and research preparation.

Capitanu required preservation of the distinction between methodological and ontological naturalism and deliberately limited the premature introduction of Intelligent Design.

This demonstrates that research and planning already contained substantive intellectual constraints before prose generation.

Part III — Chapters 5–6

By 14 June, the process involved multiple working documents together with:

• page tables;
• continuity requirements;
• source planning;
• footnotes;
• bibliography;
• provenance-sensitive planning;
• risks;
• and transition strategy.

The planning stage thus operated as a substantive bridge between accumulated research and eventual drafting.

Part IV — Chapters 7–8

By 21 June, preparatory work again included:

• target pages;
• citations;
• continuity;
• and architecture before drafting.

Capitanu added a substantive requirement concerning manufactured doubt and required conflicts of interest to be treated symmetrically.

The analysis was not to remain limited to financial interests.

It was to include non-financial forms of interest such as:

• prestige;
• reputation;
• career incentives;
• and institutional authority.

This is important evidence that pre-drafting preparation did not consist merely of organizing materials supplied by ChatGPT.

Human conceptual intervention could materially change the analytical field before prose generation.

Part V — Chapters 9–10

By 24 June, the preparation included:

• working documents;
• page targets;
• analytical dossiers;
• sources;
• architecture;
• footnote requirements;
• continuity;
• and risk analysis.

The record also preserves an explicit distinction between human planning inputs and ChatGPT proposals.

Drafting again followed validation rather than immediately following topic selection.

Part VI — Chapters 11–12

Also on 24 June, the project entered a particularly well-documented pre-drafting stage for Chapters 11–12.

The working environment identified four major input resources:

• Continuity Brief;
• Analytical Dossiers;
• Oreskes material;
• existing Manuscript Draft.

The planning stage addressed:

• chapter function;
• relevant dossiers;
• argumentative constraints;
• section architecture;
• footnotes;
• bibliography;
• risks;
• and page strategy.

The record explicitly indicates that prose had not yet been drafted at that stage.

This checkpoint is especially significant because Chapter 11 later became the representative provenance audit examined in detail in this dossier.

It provides a comparatively clear instance in which a substantial intellectual and documentary environment can be reconstructed before mature prose generation.

Part VII — Chapters 13–15

By 6 July, preparation included:

• document use;
• page targets;
• function and thesis;
• structure;
• analytical dossiers;
• continuity;
• footnote planning;
• bibliography;
• risks;
• and page strategy.

Chapter 13 additionally received more detailed planning involving:

• argument mapping;
• documentary evidence;
• dramaturgy;
• analogies;
• and section/page budgeting.

The planning model was therefore becoming increasingly granular.

Part VIII — Chapters 16–17

By 14–15 July, the planning record included:

• thesis;
• section architecture;
• dossiers;
• continuity constraints;
• footnote claims;
• bibliography;
• risks;
• page allocations;
• and search checks.

This phase also contains one of the clearest examples of substantive human intervention at the architectural level.

Capitanu rejected or superseded an initial analytical arrangement and imposed a larger five-movement architecture for Chapters 13–17, with Chapter 15 as the climax and Part VIII as the philosophical and civic resolution.

The pre-drafting stage therefore functioned as an arena of intellectual decision-making, not merely administrative preparation.

8.4 Research as a Shared Developmental Function

Research did not belong exclusively to either participant.

The record supports multiple forms of contribution.

Capitanu introduced or required particular problems, distinctions, source directions, constraints, and areas requiring investigation.

ChatGPT participated substantially in:

• identifying possible sources;
• organizing research fields;
• summarizing and comparing material;
• constructing analytical dossiers;
• identifying evidentiary gaps;
• proposing research questions;
• testing arguments against available evidence;
• and converting accumulated research into chapter-level planning.

Some source selections can be attributed specifically.

Others remain collaborative or unresolved.

The evidentiary importance lies in preserving those distinctions rather than describing the research process globally as either “human research” or “AI research.”

8.5 Pre-Drafting as Intellectual Work

The preparatory stage cannot be classified automatically as human authorship simply because it occurred before prose generation.

Some planning documents were substantially organized or generated by ChatGPT.

Their evidentiary value therefore depends on their internal provenance.

At the same time, the fact that ChatGPT generated the wording of a planning memorandum does not erase documented human contributions embodied within it.

Where Capitanu supplied:

• the research problem;
• conceptual distinctions;
• source requirements;
• chapter function;
• argumentative constraints;
• forbidden conclusions;
• placement requirements;
• or consequential corrections,

and ChatGPT transformed those inputs into a structured plan, the resulting document records both:

human intellectual direction

and

substantial AI organizational and developmental contribution.

The same analytical principle applies in the opposite direction.

Where ChatGPT independently introduced a consequential source, argument, distinction, or structural solution subsequently retained in the project, the contribution should be identified accordingly.

8.6 Development Before Text

The pre-drafting record demonstrates a fundamental distinction between:

intellectual development

and

textual generation.

Important features of the eventual manuscript sometimes existed in documented form before the prose through which they were later expressed.

These could include:

• questions;
• distinctions;
• constraints;
• source selections;
• chapter functions;
• argumentative boundaries;
• page relationships;
• or planned transitions.

This makes it possible to test a significant alternative explanation:

that the author encountered an AI-generated text first and only retrospectively recognized within it ideas later claimed as pre-existing intentions.

Where dated preparatory records demonstrably contain the relevant intellectual material before prose generation, that alternative explanation can be evaluated against contemporaneous evidence.

8.7 Pre-Drafting Did Not Eliminate AI Creativity

The existence of a detailed plan did not reduce ChatGPT to mechanical execution.

Within those plans, ChatGPT could still:

• introduce analytical developments;
• reorganize arguments;
• identify implications not explicitly specified;
• propose counterarguments;
• create formulations;
• generate analogies;
• and produce substantial original prose.

Pre-drafting therefore constrained and oriented generation without fully determining the resulting expression.

This distinction is important to the later analysis of AI contribution.

A structured intellectual environment can coexist with substantial AI analytical and expressive development.

8.8 Pre-Drafting Development Finding

The surviving documentary record supports the conclusion that research and pre-drafting organization were systematic features of the production of Why Trust Oreskes’s Plumber?.

The process was not adequately represented by repeated cycles of:

topic request → AI prose.

Across successive portions of the book, the evidence instead supports a recurring pattern of:

standing documentation → research organization → chapter-specific analysis → development of conceptual and structural architecture → human validation or redirection → authorization to draft → substantial AI and iterative development → review and revision.

The pre-drafting system was itself collaborative.

It incorporated documented human intellectual inputs and decision authority together with substantial AI analytical, organizational, and research contribution.

Its principal evidentiary significance is therefore not that it proves exclusive human authorship.

It is that it permits the developmental state of the project to be examined before prose generation and thus makes a more granular reconstruction of intellectual provenance possible.

9. Human Contributions

The surviving developmental record supports substantial human contribution throughout the production of Why Trust Oreskes’s Plumber?.

That finding does not rest upon ownership of the project, the fact that Daniel Capitanu initiated conversations with ChatGPT, the number of prompts supplied, final approval of generated material, or publication under his name.

Those facts may be relevant to the history of the project, but they are insufficient by themselves to establish substantive intellectual contribution.

The evidentiary finding rests instead upon documented participation in:

• formation of the intellectual problem;
• establishment of conceptual distinctions;
• definition of argumentative boundaries;
• selection and direction of research problems;
• imposition of constraints upon AI development;
• rejection and redirection of proposed material;
• structural decision-making;
• substantive correction;
• conceptual synthesis;
• and continuing editorial control.

The evidence does not establish that every major idea or argument in the finished book originated with Capitanu.

It does establish that his participation substantially exceeded prompting, selection, ownership, and final approval.

9.1 Intellectual Initiation

The first and most general human contribution was the initiation of the project itself.

Capitanu established the central inquiry concerning warranted scientific trust and the inspectability of the mechanisms invoked to justify that trust.

The emerging project included questions concerning:

• consensus;
• peer review;
• philosophical assumptions;
• institutional authority;
• conflicts of interest;
• historical failures;
• technological success;
• expert judgment;
• and public trust.

He also established the intended rhetorical posture of restrained or sober satire rather than anti-scientific polemic.

This initial framework did not determine every later argument.

It did establish the intellectual jurisdiction within which the book developed.

9.2 Problem Formation

A recurring human contribution throughout the project was the formulation of problems that ChatGPT subsequently developed.

Examples include:

• whether philosophical assumptions embedded in scientific practice remain open to inspection;
• whether methodological naturalism and ontological naturalism must be distinguished;
• whether exclusion of a contested inference by classification differs from refutation;
• whether Intelligent Design should function as a stress test of scientific admissibility rather than as a predetermined conclusion;
• whether procedural admissibility must remain distinct from epistemic success;
• whether science requires “philosophical charity” or inspectable trust;
• whether isolated design arguments constitute only “meteors” rather than a cumulative research program;
• and how scientific claims become consequential for citizens when mediated through institutions, private interests, policy, and governmental authority.

The evidentiary significance of these examples lies not merely in their presence in the final manuscript.

In several cases, contemporaneous records identify the human intervention before or during subsequent AI development.

9.3 Conceptual Distinctions

Capitanu repeatedly introduced or required conceptual distinctions that constrained later argument.

One of the most important was the distinction between:

methodological naturalism

and

ontological naturalism.

The project was not permitted to treat a rule governing scientific method as automatically establishing an ontological conclusion concerning what kinds of causes exist.

A related distinction concerned:

classification

versus

refutation.

An explanation could be excluded from a disciplinary category without thereby having been empirically or logically refuted.

This distinction became especially important in the Intelligent Design material.

Another governing distinction separated:

procedural admissibility

from

epistemic victory.

Allowing a claim to enter examination did not mean accepting it as true.

Likewise, criticizing a rule of exclusion did not establish the truth of the excluded claim.

These distinctions materially affected both architecture and permissible conclusions.

9.4 Intelligent Design as a Stress Test

A particularly consequential human direction concerned the role of Intelligent Design.

Capitanu required that Intelligent Design function as a stress test of admissibility, not as the predetermined verdict of the book.

The discussion was therefore constrained against two symmetrical shortcuts:

methodological exclusion → therefore Intelligent Design is false

and

criticism of methodological exclusion → therefore Intelligent Design is true.

The governing verdict was to remain suspended.

The book could examine whether methodological boundaries were argued, explicit, symmetrical, and inspectable without converting that inquiry into advocacy for a substantive conclusion not yet warranted by the evidence.

This direction became one of the most important epistemic constraints in the later manuscript.

9.5 The Suspended Verdict

The requirement of a suspended verdict was not merely a stylistic posture.

It constrained what conclusions AI-generated development was permitted to reach.

By July, surviving manuscript material expressly preserved the distinction between procedural admissibility and epistemic victory.

Neither methodological naturalism nor Intelligent Design was to receive automatic epistemic privilege merely by classification rule.

The human contribution here consisted not only in selecting among existing outputs, but in defining in advance the permissible epistemic position of the developing argument.

9.6 “The Right to Be Wrong”

On 30 June 2026, Capitanu introduced the principle that science possesses a conditional right to be wrong, provided that error remains visible to correction.

He extended this principle to controversial claims, including those involving Intelligent Design.

This intervention changed the structure of the admissibility problem.

The issue became not merely whether disciplinary boundaries can be inspected, but whether a culture committed to fallibility can justify excluding a sufficiently serious claim from the arena in which error can become visible, criticized, and corrected.

ChatGPT subsequently developed and formulated the principle.

The provenance distinction is therefore important:

governing concept — HUMAN

development and formulation — AI / ITERATIVE

The contribution cannot responsibly be attributed wholly to either participant.

9.7 Philosophical Charity and Inspectable Trust

Capitanu later raised the question whether science requires the kind of philosophical charity sometimes requested for it or whether warranted trust instead requires inspection.

ChatGPT developed this intervention into the proposed section “Does Science Need Charity, or Inspection?” and further developed the concept of inspectable trust.

ChatGPT initially proposed placing the material in Chapter 17.

Capitanu relocated it to Chapter 15 in order to preserve Chapter 17 as the concluding synthesis.

This sequence contains multiple distinct human contributions:

• the initiating intellectual question;
• the decision that inspection rather than charity should frame the problem;
• and the final architectural placement.

The resulting formulation nevertheless reflects substantial AI development.

9.8 From “Meteors” to a Research Program

Capitanu introduced the “meteors” analogy to characterize isolated Intelligent Design observations or arguments that might be interesting individually but would not, by themselves, constitute a mature scientific research program.

He then asked what Intelligent Design would have to accomplish in order to become scientifically competitive.

ChatGPT transformed this into a developed framework concerned with requirements such as:

• discriminating predictions;
• falsifiability;
• cumulative research performance;
• comparative explanatory performance;
• and the avoidance of mere reliance upon explanatory gaps.

This is another relatively clear case of:

human conceptual initiation → substantial AI analytical development → manuscript integration.

The human contribution lies in formulating the problem and demanding a higher standard than isolated argumentative successes.

9.9 Institutional Mediation and Civic Consequences

Capitanu introduced COVID-19 as a case through which to examine the path by which scientific findings become consequential for citizens.

The concern was not restricted to science as an epistemic practice.

It included the movement of claims through:

• expert bodies;
• institutional interpretation;
• private interests;
• governmental decision-making;
• policy;
• and public authority.

He extended the concern prospectively to artificial intelligence, biotechnology, biohacking, and other emerging technologies.

ChatGPT then systematized this intervention into a more explicit sequence distinguishing evidence, interpretation, institutional judgment, private interests, policy choice, and public authority.

The human contribution here broadened the jurisdiction of the book from epistemic trust alone toward the civic consequences of mediated expertise.

9.10 Research Direction and Constraints

Capitanu repeatedly contributed to research direction by determining:

• which questions required investigation;
• which controversies should be delayed;
• which distinctions must be preserved;
• where symmetrical treatment was required;
• which forms of evidence were insufficient;
• and when additional research was necessary before drafting or conclusion.

One documented example concerns conflicts of interest.

The treatment was not to remain limited to industry funding or other financial interests.

Capitanu required the analysis to include non-financial interests such as:

• prestige;
• reputation;
• career incentives;
• and institutional authority.

This broadened the analytical framework before subsequent prose generation.

9.11 Structural Intervention

Human contribution also operated at the architectural level.

Capitanu:

• accepted but subsequently modified the initial seventeen-chapter architecture;
• limited the scope of early chapters;
• determined when controversial topics could enter;
• relocated material when proposed placement conflicted with the intended argument;
• preserved Chapter 17 as concluding synthesis;
• determined relationships between Chapters 11 and 12;
• and later imposed a broader architecture in which Chapters 13–17 functioned as a five-movement sequence with Chapter 15 as climax and the final part as philosophical and civic resolution.

These are consequential structural decisions rather than merely administrative approvals.

9.12 Rejection and Redirection

Human contribution is especially visible where ChatGPT proposals were not accepted.

Capitanu could:

• reject an analytical route;
• narrow an argument;
• prevent premature treatment of a controversy;
• require greater fairness to an opposing position;
• preserve a suspended verdict;
• move material between chapters;
• require additional nuance;
• demand further evidence;
• or supersede an existing architecture.

Rejection is evidentially important because it reveals the existence of governing intellectual criteria independent of the availability of generated output.

9.13 Substantive Correction

The record also identifies repeated human interventions intended to prevent oversimplification.

This included corrections concerning:

• the treatment of opposing positions;
• the relationship between methodology and ontology;
• the difference between admissibility and truth;
• the evidentiary burden placed upon contested explanations;
• and the limits of what particular examples could establish.

Substantive correction differs from proofreading.

It alters the reasoning, evidentiary warrant, conceptual balance, or permissible conclusion of the work.

9.14 Selection and Integration

Capitanu ultimately determined which material entered the manuscript and where it entered.

Selection, however, is not treated in this dossier as automatic authorship of selected AI-generated expression.

Its significance depends upon the intellectual judgment embodied in the act.

Selection becomes more probative where it forms part of a documented sequence involving:

• prior problem formation;
• alternative analysis;
• rejection;
• restructuring;
• revision;
• and integration into a larger argumentative architecture.

9.15 Editorial and Canonical Control

The later development of the manuscript provides strong evidence of continuing human editorial authority.

Capitanu determined:

• whether drafting could proceed;
• whether arguments exceeded their evidentiary warrant;
• whether structures were retained or replaced;
• whether sections required further work;
• when chapters were acceptable;
• how the manuscript was canonicalized;
• and which production versions were authorized.

This authority continued through later print, EPUB, and audiobook production.

Publication control alone would not establish human intellectual authorship.

Its significance in the present record lies in its continuity with earlier documented intellectual direction.

9.16 Human-Contribution Finding

The evidence does not support a model in which Capitanu merely:

• selected a topic;
• issued high-level prompts;
• chose among finished AI outputs;
• corrected superficial defects;
• and authorized publication.

His documented contributions include substantive activity in:

problem formation, conceptual distinction, research direction, argumentative constraint, structural decision-making, rejection, redirection, substantive correction, intellectual synthesis, integration, and editorial control.

At the same time, this finding does not justify attributing every major idea, argument, structure, or expression in the work to him.

The appropriate conclusion is narrower and better supported:

Daniel Capitanu made substantial and continuing human intellectual contributions to the development of Why Trust Oreskes’s Plumber?, including contributions before, during, and after AI generation.

That finding is compatible with the simultaneously substantial AI contribution examined in the next section.

10. AI Contributions

The contribution of OpenAI’s ChatGPT to Why Trust Oreskes’s Plumber? was substantial, sustained, and functionally diverse.

The evidence does not support describing ChatGPT as merely:

• a proofreader;
• copy editor;
• research utility;
• spelling or grammar tool;
• formatting assistant;
• or occasional brainstorming aid.

Its participation extended into research, architecture, analytical development, argumentation, drafting, redrafting, critical testing, continuity management, and editorial development.

The documentary record therefore supports a strong finding:

the AI contribution was substantive and cannot responsibly be minimized.

At the same time, the significance of particular AI activities differs.

Some functions resemble conventional research or literary assistance.

Others resemble developmental editing or intellectual collaboration.

Still others involved direct generation of substantial expressive material.

These functions must be distinguished rather than collapsed into a single category.

10.1 Assistive Functions

ChatGPT performed many activities with recognizable analogues in conventional human research, literary, scholarly, or editorial assistance.

These included:

• locating and organizing potentially relevant information;
• identifying source requirements;
• summarizing and comparing source material;
• preparing research-oriented working documents;
• assisting with source verification;
• assisting with bibliography and footnote preparation;
• maintaining continuity across a very large manuscript;
• identifying contradictions and redundancies;
• monitoring version consistency;
• identifying structural problems;
• supporting production decisions;
• and assisting with editorial diagnosis.

These functions were often extensive.

Their extent is evidence of substantial AI participation.

Their existence does not, by itself, determine authorship.

10.2 Research Organization

ChatGPT participated substantially in transforming broad research fields into usable working structures.

It helped organize material concerning:

• scientific trust;
• consensus;
• philosophy of science;
• methodological naturalism;
• peer review;
• conflicts of interest;
• historical cases;
• Intelligent Design;
• institutional authority;
• contemporary research problems;
• and the relationship between scientific evidence and public policy.

It also participated in maintaining Analytical Dossiers and other preparatory materials.

The result was not merely retrieval of isolated facts.

Research material was organized according to the argumentative needs of chapters and sections.

10.3 Source Identification and Verification

ChatGPT assisted in identifying potential sources and in checking or organizing source-related material.

This included work relevant to:

• quotations;
• historical claims;
• scientific cases;
• footnotes;
• bibliography;
• and supporting documentation.

Source assistance should not be confused with the independent establishment of evidentiary truth.

The project still required judgment regarding:

• relevance;
• reliability;
• interpretation;
• sufficiency;
• and ultimate inclusion.

Nevertheless, the source-related contribution was materially greater than incidental search assistance.

10.4 Book-Level Architecture

One of the earliest consequential AI contributions was the proposal of the seventeen-chapter architecture.

ChatGPT transformed the developing field of questions into a sustained sequence involving:

• Oreskes;
• consensus;
• philosophical assumptions;
• historical and institutional problems;
• peer review;
• technological success;
• contested inference;
• contemporary scientific problems;
• and public trust.

This architecture was accepted as the working structure and survived in broad form through the later manuscript, although extensively developed and modified.

It constitutes clear evidence that ChatGPT contributed materially at the level of overall book architecture.

10.5 Chapter and Section Architecture

AI contribution continued below the book level.

ChatGPT proposed:

• chapter organizations;
• section sequences;
• internal argumentative structures;
• transitions;
• subdivisions;
• placement options;
• and drafting architectures.

Some were accepted.

Some were modified.

Some were relocated.

Some were rejected.

The fact that human decision authority remained operative does not erase the architectural contribution.

10.6 Converting Constraints Into Executable Plans

One recurring AI function was the transformation of human-defined objectives, constraints, and research materials into detailed drafting plans.

A human instruction might establish:

• the problem;
• limits;
• required distinctions;
• source requirements;
• prohibited shortcuts;
• chapter function;
• and desired relationship to the rest of the book.

ChatGPT could then convert those elements into:

• an analytical memorandum;
• section architecture;
• argumentative sequence;
• research agenda;
• page strategy;
• and executable drafting plan.

This role lies between conventional assistance and direct textual generation.

It is appropriately classified as developmental.

10.7 Analytical Development

ChatGPT repeatedly expanded intellectual material beyond what had been exhaustively specified in preceding human instructions.

This included:

• developing implications of concepts;
• distinguishing competing interpretations;
• identifying hidden assumptions;
• connecting previously separate ideas;
• creating analytical categories;
• expanding argumentative consequences;
• and transforming broad questions into structured intellectual frameworks.

The “meteors” case illustrates this function.

Capitanu introduced the question whether isolated Intelligent Design arguments could amount to a serious research program.

ChatGPT developed that question into a more systematic framework involving discriminating predictions, falsifiability, cumulative research performance, and comparative performance against alternative explanations.

This was not simply transcription of a fully specified human answer.

It was substantial analytical development.

10.8 Development of “Inspectable Trust”

Another example concerns philosophical charity and inspectable trust.

Capitanu introduced the question whether science requires philosophical charity or inspection.

ChatGPT developed the intervention into:

• a proposed section;
• a more elaborate conceptual treatment;
• and the formulation of inspectable trust.

The phrase later became central to the book’s conceptual vocabulary.

The contribution is therefore appropriately distinguished as:

human initiating problem → substantial AI conceptual and expressive development.

10.9 Argument Generation

ChatGPT contributed to the generation and expansion of arguments.

Its role included:

• identifying argumentative pathways;
• constructing intermediate steps;
• articulating implications;
• proposing objections;
• generating responses;
• and testing whether conclusions followed from premises.

Some arguments were substantially generated or materially developed by ChatGPT rather than being fully contained in the preceding human input.

This fact is central to the provenance record.

The dossier therefore deliberately avoids the claim that all “core arguments” originated with the human author.

10.10 Counterarguments and Adversarial Testing

ChatGPT also performed a sustained critical-testing role.

It generated:

• counterarguments;
• alternative interpretations;
• objections;
• evidentiary challenges;
• and possible weaknesses in developing claims.

This adversarial function was used both during development of the book and later during construction of the provenance record itself.

The significance of this contribution lies in its capacity to alter the argument rather than merely polish its presentation.

10.11 Structural Diagnosis and Restructuring

ChatGPT identified:

• redundancy;
• discontinuity;
• misplaced material;
• insufficient transitions;
• overlong sections;
• architectural imbalance;
• and possible alternative arrangements.

It also proposed reorganizations.

Some proposals were retained.

Others were overridden by human architectural decisions.

The appropriate provenance account recognizes both:

AI structural initiative

and

human authority to accept, modify, or reject that initiative.

10.12 Drafting and Redrafting

ChatGPT generated substantial prose intended for possible incorporation into the manuscript.

This included:

• sentences;
• paragraphs;
• sections;
• transitions;
• explanatory passages;
• satirical passages;
• argumentative prose;
• and longer chapter material.

It also participated in redrafting existing material.

This contribution cannot responsibly be redescribed merely as “language assistance.”

Where ChatGPT produced expressive material, the record identifies the function as generative/expressive.

The existence of human constraints, intellectual inputs, or later revision does not alter the immediate fact of AI textual generation.

10.13 Formulations and Analogies

ChatGPT contributed formulations, rhetorical devices, metaphors, and analogies.

Some arose directly from human-originating concepts.

Others were developed through iterative exchange.

Some were introduced by ChatGPT and subsequently retained.

The provenance of a memorable formulation therefore must not be inferred merely from the origin of the underlying concept.

The “right to be wrong” case provides a clear example.

The governing principle originated with Capitanu.

ChatGPT subsequently developed and formulated it into expressive language.

The concept and the wording consequently have different provenance.

10.14 Chapter 11 Analytical Development

The reconstruction of Chapter 11 provides particularly strong evidence of substantive AI participation.

Within a chapter constrained by prior human decisions, ChatGPT substantially developed material concerning:

• examination before dismissal;
• minimum seriousness;
• fallible scientific borders;
• refutation rather than exclusion by classification;
• procedural admissibility versus epistemic victory;
• fertile error;
• curricular distinctions;
• symmetry;
• and restrained plumber formulations.

The AI contribution here was not merely mechanical execution of exhaustively specified instructions.

It involved substantive analytical and expressive development within an existing intellectual jurisdiction.

10.15 Continuity Management

The scale of the manuscript made continuity itself a substantial production problem.

ChatGPT participated in maintaining awareness of:

• earlier arguments;
• terminology;
• chapter relationships;
• unresolved questions;
• recurring distinctions;
• duplication;
• and internal consistency.

This function was especially important because the book was developed over many sessions and through multiple working documents.

Continuity management resembles an assistive or editorial function, but its scale materially influenced the possibility of sustaining the project.

10.16 Working Documentation

ChatGPT participated in preparing and maintaining:

• Continuity Briefs;
• Analytical Dossiers;
• planning memoranda;
• chapter architectures;
• section plans;
• research summaries;
• provenance-sensitive notes;
• and other working documents.

Some of these records were substantially generated by ChatGPT.

Their AI origin does not deprive them of evidentiary value, but it requires the dossier to distinguish between:

• AI wording and organization;
• human inputs embodied within them;
• AI-originating developments;
• and iterative decisions.

10.17 Editorial Revision

ChatGPT also participated in revision after initial drafting.

This included:

• rewriting;
• compression;
• expansion;
• clarification;
• restructuring;
• transition work;
• identification of inconsistencies;
• and improvement of continuity.

Some revisions were initiated in response to specific human criticism.

Others were proposed through AI analysis.

The resulting text therefore often reflects multiple developmental stages rather than a single act of generation.

10.18 AI Initiative

The record contains evidence that ChatGPT sometimes contributed material not exhaustively contained in preceding human instructions.

This included:

• architectural proposals;
• analytical distinctions;
• argumentative expansions;
• counterarguments;
• structural solutions;
• formulations;
• and expressive material.

Some of these contributions survived into the developing manuscript.

This is important because it excludes a model in which ChatGPT functioned solely as a deterministic execution mechanism for complete human specifications.

AI participation included genuine developmental initiative within the constraints of the system and project.

10.19 Limits of AI Authority

Substantial contribution did not entail independent decision authority over the completed work.

ChatGPT could propose, develop, reorganize, critique, and generate.

It did not possess authority to determine:

• the ultimate purpose of the project;
• which governing constraints should prevail;
• which conclusions were permissible;
• which material entered the canonical manuscript;
• whether a proposed architecture was accepted;
• when a chapter was complete;
• or when a publication master was authorized.

This distinction does not minimize AI contribution.

It separates contribution from decision authority.

10.20 Functional Classification

The evidence supports at least three categories of ChatGPT contribution.

ASSISTIVE

Functions such as:

• research organization;
• source assistance;
• continuity management;
• fact-checking;
• bibliographic work;
• editorial diagnosis;
• and analytical testing.

These can be substantial without necessarily involving authorship of expression.

DEVELOPMENTAL

Functions such as:

• proposing structures;
• expanding concepts;
• generating counterarguments;
• reorganizing argumentative sequences;
• identifying implications;
• and converting author-defined intellectual material into developed plans.

These functions occupy a more complex intermediate category.

GENERATIVE / EXPRESSIVE

Functions such as:

• producing sentences;
• paragraphs;
• sections;
• analogies;
• formulations;
• transitions;
• and other potentially publishable prose.

These contributions must be acknowledged directly as AI-generated expression where the evidence supports that description.

10.21 Functionally Plural Role

No single conventional label captures the complete role performed by ChatGPT.

At different stages, the system functioned analogously to combinations of:

• research assistant;
• analytical assistant;
• developmental editor;
• literary assistant;
• critical interlocutor;
• structural planner;
• and substantial drafting instrument or collaborator.

None of these analogies should erase the others.

Calling ChatGPT only a “research assistant” would understate its analytical and expressive contribution.

Calling it only a “writer” would obscure the extent to which much of its work occurred within human-originating objectives, constraints, research problems, and final decision authority.

10.22 AI-Contribution Finding

The surviving evidence supports the following conclusion:

ChatGPT made substantial contributions to the research, organization, architecture, conceptual and argumentative development, critical testing, drafting, redrafting, revision, continuity management, and expressive realization of Why Trust Oreskes’s Plumber?.

Some of these functions were assistive.

Some were developmental.

Some involved direct generation of potentially publishable expression.

The contribution was sufficiently extensive that descriptions such as:

“AI-assisted proofreading,”

“AI copyediting,”

or

“AI used only for research”

would materially misrepresent the production record.

At the same time, substantial AI contribution does not, by itself, establish legal authorship, co-authorship, autonomous creation of the work, or authority over the final manuscript.

The relevant finding is narrower:

AI participation was substantial, consequential, and integrated into the intellectual and expressive development of the work.

11. Iterative Collaboration

The development of Why Trust Oreskes’s Plumber? was not organized around a stable division in which one participant supplied ideas and the other merely supplied language.

Nor did the process consist of a single transfer of intellectual control from human to machine or machine to human.

The surviving record instead supports a recurring pattern of iterative collaboration in which human and AI contributions entered at different moments, altered one another, and were repeatedly subjected to evaluation, redirection, expansion, revision, and integration.

For purposes of this record, iterative collaboration is a descriptive term.

It does not imply legal co-authorship.

It identifies a production process in which the developmental history of some material cannot be represented accurately by assigning the entire contribution to only one participant.

11.1 Why Binary Attribution Is Often Inadequate

Some contributions can be attributed relatively cleanly.

A documented human intervention may introduce a concept.

A documented AI response may introduce a new formulation.

A human decision may relocate a section.

An AI draft may supply a particular passage.

Other developments are more complex.

A concept may:

1. originate with Capitanu;
2. be expanded by ChatGPT;
3. be criticized or constrained by Capitanu;
4. be reorganized by ChatGPT;
5. receive additional human qualifications;
6. be redrafted by ChatGPT;
7. and enter the canonical manuscript after further human selection and integration.

Assigning the final result wholly to either participant would erase relevant parts of the developmental sequence.

The appropriate analytical task is therefore not always:

Who created this?

It may instead be:

How did this develop, and what contribution did each participant make at successive stages?

11.2 The Recursive Developmental Pattern

Across the project, the record supports a recurring sequence of interaction:

human problem or constraint → AI proposal or development → human evaluation → acceptance, correction, rejection, or redirection → further AI development → additional human intervention → integration.

The sequence could repeat multiple times within the same chapter or conceptual problem.

In some cases, ChatGPT introduced a structure or formulation that Capitanu accepted.

In others, AI proposals were substantially altered or rejected.

In still others, a human-originating concept acquired its mature analytical form primarily through AI development.

This recursive pattern is central to understanding the production process.

11.3 Collaboration Before Drafting

Iterative collaboration often began before prose generation.

The pre-drafting environment itself could contain contributions from both participants.

Capitanu might supply:

• the research problem;
• relevant source material;
• conceptual distinctions;
• chapter function;
• argumentative constraints;
• prohibited conclusions;
• placement requirements;
• or corrections to an existing plan.

ChatGPT might then:

• organize those inputs;
• identify implications;
• propose an architecture;
• expand the research field;
• identify risks;
• generate counterarguments;
• or convert the material into a detailed working plan.

That plan could then be accepted, modified, or rejected.

Pre-drafting documentation is therefore neither presumptively human nor presumptively AI.

Its provenance must be reconstructed internally.

11.4 Collaboration During Generation

Once drafting began, the interaction could become more complex.

ChatGPT generated substantial prose and analytical development.

Capitanu continued to intervene in:

• permissible conclusions;
• conceptual distinctions;
• fairness to opposing positions;
• placement;
• argumentative balance;
• evidentiary warrant;
• and relationship to the larger architecture.

Generated prose therefore did not end the intellectual process.

It often became material for further intellectual decision-making.

11.5 Collaboration After Generation

Post-generation activity included more than proofreading.

The record identifies:

• substantive correction;
• restructuring;
• relocation;
• requests for additional research;
• sharpening or narrowing of claims;
• rejection of overstatement;
• modification of argumentative sequence;
• integration with surrounding chapters;
• and redrafting.

Accordingly, a passage initially generated by ChatGPT may have entered a larger developmental process before reaching canonical form.

The initial generator of the wording and the final intellectual provenance of the published unit are therefore analytically distinct questions.

11.6 Concept and Development as Separate Contributions

The “right to be wrong” material provides a particularly clear example.

Capitanu introduced the governing concept that science possesses a conditional right to be wrong insofar as its errors remain visible to correction.

ChatGPT then developed the idea analytically and integrated it into the Chapter 11 argument.

The surviving record therefore supports:

GOVERNING CONCEPT — HUMAN

ANALYTICAL DEVELOPMENT — AI / ITERATIVE

EXPRESSIVE FORMULATION — AI

The example demonstrates why neither of the following statements would be adequate:

“ChatGPT created the idea.”

or

“Capitanu created the entire resulting argument.”

The first would erase documented human conceptual origin.

The second would erase substantial AI analytical and expressive development.

11.7 Human Constraint and AI Expansion

Another recurring pattern involved human establishment of a governing boundary followed by AI development within that boundary.

For Chapter 11, Capitanu required that Intelligent Design function as an admissibility stress test rather than as a predetermined conclusion.

He also required preservation of the distinction between:

admissibility

and

epistemic victory.

Within those constraints, ChatGPT developed further material concerning:

• examination before dismissal;
• minimum seriousness;
• fallible scientific borders;
• classification versus refutation;
• productive error;
• curricular distinctions;
• symmetrical inspection;
• and related formulations.

The resulting material was therefore neither fully specified by the human input nor independent of it.

It represents substantive AI development within a human-established intellectual jurisdiction.

11.8 AI Proposal and Human Redirection

The collaboration also worked in the opposite direction.

ChatGPT could introduce:

• an architecture;
• a placement;
• a distinction;
• a transition;
• a formulation;
• or a line of argument.

Capitanu could then:

• accept it;
• relocate it;
• narrow it;
• broaden it;
• reject it;
• or integrate it differently.

The “inspectable trust” material is representative.

Capitanu raised the initiating philosophical question.

ChatGPT developed the concept and proposed placement in Chapter 17.

Capitanu relocated it to Chapter 15 to preserve the function of Chapter 17 as concluding synthesis.

The final result therefore reflects:

human initiation → AI conceptual development → AI placement proposal → human architectural redirection → integrated canonical function.

11.9 Architecture as Iterative Collaboration

The initial seventeen-chapter architecture was a substantial ChatGPT contribution.

It did not remain unchanged.

The architecture was repeatedly developed through:

• human constraints;
• AI planning;
• human rejection or relocation;
• new AI proposals;
• and later higher-level human restructuring.

The final architecture therefore has mixed provenance.

Its macro-form reflects a consequential early AI proposal.

Its later functions and relationships reflect extensive iterative development under continuing human direction.

11.10 Research as Iterative Collaboration

Research followed a similar pattern.

Capitanu often determined:

• what problem required research;
• what distinctions were relevant;
• what claims required verification;
• and what conclusions the evidence would have to support.

ChatGPT contributed by:

• locating or suggesting sources;
• organizing materials;
• comparing claims;
• identifying evidentiary gaps;
• constructing dossiers;
• and transforming research into analytical structures.

Research provenance therefore cannot be represented globally as either human or AI.

Different research acts had different origins and functions.

11.11 Argumentative Iteration

Arguments could emerge through repeated cycles.

A typical sequence could involve:

1. human identification of a problem;
2. AI generation of a possible argumentative route;
3. human objection to an implication;
4. AI reformulation;
5. human requirement for symmetry or qualification;
6. AI generation of a more balanced structure;
7. later integration into the chapter.

The final argument may therefore contain multiple identifiable layers of contribution.

This is one reason the dossier avoids artificial percentage allocations.

A numerical division of “authorship” would obscure rather than clarify the developmental history.

11.12 Asymmetry Within Collaboration

Iterative collaboration does not imply equality of function or authority.

The participants did not perform identical roles.

ChatGPT could generate substantial analytical and expressive material.

Capitanu retained authority to determine:

• the governing purpose of the project;
• permissible conclusions;
• which arguments required further work;
• which structures were retained;
• which text entered the manuscript;
• and when a publication state became canonical.

This asymmetry is analytically important.

Collaboration does not require identical decision rights.

11.13 Collaboration Does Not Erase Provenance

Calling the process iterative does not mean that individual contributions become untraceable.

Where the record permits, this dossier continues to identify:

• human-originating concepts;
• AI-originating developments;
• AI-generated formulations;
• human rejections;
• AI structural proposals;
• human placement decisions;
• and other discrete contributions.

The label ITERATIVE is used only when the relevant unit is genuinely the product of successive interaction.

It is not a substitute for granular attribution where granular attribution is possible.

11.14 Representative Chapter 11 Pattern

The Chapter 11 reconstruction provides the clearest concentrated example of the iterative model.

The surviving evidence supports the following developmental sequence:

human intellectual problem and governing constraints

→

collaborative pre-drafting development

→

substantial AI analytical and textual generation

→

continuing human conceptual intervention and editorial control

→

collaboratively developed canonical chapter.

Neither of the simplified alternatives fits the record adequately.

The chapter was not independently devised and written by ChatGPT from a high-level human topic.

Nor was it a substantially human-written chapter merely polished by AI.

The evidence supports consequential contribution on both sides across different developmental stages.

11.15 Iterative-Collaboration Finding

The record therefore supports iterative collaboration as the most accurate general description of the developmental mechanism.

The term captures a process in which:

• human intellectual inputs repeatedly shaped AI development;
• AI contributions materially expanded the intellectual and expressive possibilities available to the author;
• generated material became subject to human criticism, redirection, and integration;
• and subsequent AI development responded to those interventions.

This does not determine legal authorship.

It establishes the developmental structure of the collaboration.

The resulting book is best understood not as the output of a single transfer of instruction, but as the product of a sustained recursive production process in which intellectual agency, analytical development, expression, and decision authority were distributed differently across successive stages.

12. Drafting and Expressive Contribution

The provenance of ideas and the provenance of expression must be examined separately.

A concept may originate with one participant while the wording through which it appears in the manuscript is generated by the other.

Likewise, AI-generated prose may embody:

• human-originating concepts;
• human-defined constraints;
• collaboratively developed arguments;
• AI-originating analytical developments;
• or combinations of all of these.

For this reason, the fact that a passage was initially generated by ChatGPT does not determine the provenance of every intellectual element contained within it.

Conversely, the fact that Capitanu originated a concept does not establish that he generated the later expression through which that concept was articulated.

The record therefore treats drafting and expressive contribution as a distinct provenance dimension.

12.1 Direct AI Prose Generation

ChatGPT generated substantial original prose intended for possible incorporation into Why Trust Oreskes’s Plumber?.

This occurred from the early stages of the project.

By late May, ChatGPT was already producing substantial prose for:

• the Prologue;
• the early chapters;
• and later, successive sections and chapters across the manuscript.

Its participation was therefore not limited to:

• proofreading;
• stylistic correction;
• copyediting;
• or restructuring prose written independently by Capitanu.

It was a direct participant in textual generation.

12.2 Generative/Expressive Functions

For purposes of this record, generative/expressive functions include the production of:

• sentences;
• paragraphs;
• sections;
• transitions;
• analogies;
• formulations;
• satirical language;
• explanatory prose;
• argumentative prose;
• and other text capable of appearing in the published work.

These contributions must be acknowledged directly.

They should not be redescribed euphemistically as “language assistance” where the evidence shows actual AI generation of publishable expression.

12.3 Drafting Within a Pre-Existing Intellectual Environment

AI drafting usually occurred within an established developmental environment.

For example, by the time substantial Chapter 11 generation was underway, ChatGPT already had access to an intellectual and documentary context containing:

• the book-level problem of warranted and inspectable trust;
• existing manuscript continuity;
• the distinction between methodological and ontological naturalism;
• accumulated research on methodological naturalism and demarcation;
• classification versus substantive refutation;
• Intelligent Design as an admissibility stress test;
• the rule that admissibility must not be confused with epistemic victory;
• the division of labor between Chapters 11 and 12;
• the requirement of a suspended verdict;
• and the developing principle that science must retain a capacity to be wrong and correct itself.

AI-generated prose must therefore be interpreted against that prior intellectual context.

The expressive contribution was substantial.

It was not context-free.

12.4 Expression Does Not Determine Conceptual Origin

The strongest illustration is the “right to be wrong” material.

Capitanu introduced the governing concept.

ChatGPT developed the argument.

One resulting formulation expressed the distinction between the right to be wrong and the right to be believed, connecting error with the right to examination before dismissal.

The evidentiary conclusion is therefore:

conceptual origin — HUMAN

analytical development — AI / ITERATIVE

expressive formulation — AI

This is a concrete example of why sentence-level origin cannot be used as a shortcut for intellectual provenance.

12.5 AI-Originating Expression

The record also contains instances in which the expressive contribution itself can be attributed specifically to ChatGPT.

In the Chapter 11 development, ChatGPT generated plumber-related formulations translating the procedural problem into the book’s established satirical language.

One documented formulation was accepted during development and linked the refusal to test a leak with failure to establish that the plumber was wrong.

The evidentiary classification is:

AI-GENERATED EXPRESSION + HUMAN SELECTION WITHIN AN ESTABLISHED BOOK-LEVEL METAPHOR.

This example matters because the plumber metaphor as a governing device predated the particular sentence.

The specific expression and the underlying metaphor therefore have different provenance.

12.6 Human-Defined Voice and AI-Generated Language

The production process also demonstrates that voice and immediate textual generation can be separated.

Capitanu established broad rhetorical requirements, including:

• sober satire;
• fairness to opposing positions;
• avoidance of anti-scientific polemic;
• restraint;
• inspectability;
• and preservation of epistemic distinctions.

ChatGPT generated large quantities of prose within those constraints.

Thus, a passage may exhibit a project-level rhetorical identity shaped by human direction while still being AI-generated at the sentence level.

This distinction is important for any later copyright or authorship analysis.

12.7 Redrafting

ChatGPT also participated substantially in redrafting.

Redrafting involved more than correcting grammar.

It could include:

• reformulation;
• expansion;
• compression;
• reorganization;
• sharpening of argument;
• replacement of weak transitions;
• restructuring of paragraphs;
• and adaptation of material to changed chapter architecture.

A redrafted passage may therefore have a layered textual provenance.

Its first version may have been AI-generated.

Its later structure may reflect human criticism.

Its revised wording may again have been generated by ChatGPT.

A later human intervention may then alter or select the final form.

12.8 Human Textual Intervention

The surviving record supports substantial human intellectual and editorial intervention, but it does not always permit precise quantification of sentence-level human rewriting.

The Chapter 11 audit explicitly leaves unresolved:

• how much prose was directly rewritten by Capitanu;
• which AI-generated passages were rejected;
• and how extensively the earliest generated text changed before reaching canonical form.

These matters should not be filled retrospectively.

Where the documentary record does not establish sentence-level human rewriting, this record does not claim it.

This limitation is important because the dossier seeks accurate provenance rather than maximal human attribution.

12.9 Rejection as Part of Expressive Provenance

Generated prose did not become part of the book merely because ChatGPT produced it.

The author could:

• reject a formulation;
• request another version;
• alter the function of a section;
• relocate material;
• require a different emphasis;
• or exclude the text entirely.

Rejection is therefore part of textual history.

At the same time, rejection of alternatives does not transform the retained AI-generated wording into human-generated wording.

Selection and generation remain separate dimensions.

12.10 Selection and Retention

Human selection is relevant to the history of the published text.

It determines which generated material became part of the developing manuscript.

However, this dossier does not adopt the proposition that selecting a sentence among AI outputs automatically makes the selector the generator or author of that sentence.

The appropriate description is narrower:

AI-generated expression may enter the manuscript through consequential human selection and integration.

Both facts can be simultaneously true.

12.11 Expression Through Iteration

Some final passages cannot be represented adequately as a single act of generation.

They may have developed through:

AI generation → human criticism → AI redraft → human restructuring → AI reformulation → human selection.

In such cases, the immediate provenance of particular words may remain AI while the broader expressive unit reflects iterative development.

The appropriate level of attribution depends upon the unit being examined.

12.12 Substantial Analytical Development During Drafting

Drafting itself also generated intellectual developments.

ChatGPT did not merely convert pre-existing concepts into sentences.

During generation, it sometimes introduced:

• new analytical distinctions;
• implications;
• conceptual extensions;
• symmetry principles;
• counterarguments;
• thresholds;
• and new argumentative connections.

In Chapter 11, documented developments included:

• examination before dismissal;
• minimum seriousness;
• fallible scientific borders;
• refutation versus expulsion by classification;
• productive error;
• curricular distinctions;
• symmetrical inspection;
• and further plumber formulations.

These contributions show that the boundary between “analysis” and “writing” was porous.

Text generation sometimes became a site of intellectual development.

12.13 Drafting as a Developmental Instrument

The project therefore did not always proceed according to:

fully complete idea → neutral transcription into prose.

Drafting could itself expose:

• weaknesses;
• hidden assumptions;
• missing transitions;
• contradictory implications;
• and unexplored consequences.

The act of generating prose could produce material that was then evaluated intellectually.

This is another reason textual and conceptual provenance must be analyzed separately but not artificially isolated from one another.

12.14 The Literary-Assistant Analogy and Its Limit

Some AI functions have strong analogues in conventional human literary assistance.

A research assistant, developmental editor, or editorial assistant may perform extensive work without generating the work’s expressive core.

That analogy becomes insufficient when ChatGPT generates substantial potentially publishable passages or original expressive formulations.

The record therefore refuses to use literary assistant as a blanket description of all AI activity.

Where expression was AI-generated, it is identified as such.

12.15 Drafting Versus Authorship

The fact that ChatGPT generated substantial prose does not by itself settle legal authorship.

Likewise, human control of the larger project does not automatically convert AI-generated expression into human-generated expression.

The present record does not attempt to resolve that legal question.

It preserves the more basic factual distinction:

who generated expression

is not necessarily the same question as

who originated the concept

or

who controlled the final work.

12.16 Textual Provenance and Canonical Form

The canonical manuscript represents the end state of a developmental process.

It establishes what survived.

It does not by itself reveal how every surviving sentence originated.

For some passages, conversational records allow relatively direct attribution.

For others, only mixed or partial provenance can be reconstructed.

For still others, exact passage-level history remains unresolved.

Canonical presence should therefore not be mistaken for proof of original textual provenance.

12.17 Expression as a Separate Evidentiary Layer

The production history is best understood through at least three analytically distinct layers:

INTELLECTUAL PROVENANCE

Where did the concept, question, argument, or governing constraint originate?

DEVELOPMENTAL PROVENANCE

How was that intellectual material expanded, structured, tested, qualified, or transformed?

EXPRESSIVE PROVENANCE

Who or what generated the wording through which the developed material was expressed?

A fourth layer remains distinct:

EDITORIAL DISPOSITION

Who accepted, rejected, revised, relocated, integrated, and ultimately authorized the material?

These layers may have different answers for the same passage.

12.18 Representative Provenance Pattern

A representative pattern found in the project is:

human-originating concept

→

AI analytical expansion

→

AI-generated formulation

→

human evaluation and constraint

→

AI redrafting

→

human selection and integration.

Another pattern is:

AI-originating structural or argumentative proposal

→

human redirection

→

AI reformulation

→

iterative development

→

canonical incorporation.

Neither pattern can be captured accurately by a single global label such as “human-written” or “AI-written.”

12.19 Drafting and Expressive Contribution Finding

The evidence supports a strong finding that ChatGPT generated substantial expressive material for Why Trust Oreskes’s Plumber?.

That fact must be stated without euphemism.

At the same time, textual generation does not determine the provenance of every intellectual component embodied in the generated text.

The record therefore distinguishes:

• human-originating concepts expressed in AI-generated prose;
• AI-developed arguments expressed in AI-generated prose;
• iteratively developed material expressed through AI-generated formulations;
• human-modified or redirected AI-generated material;
• and passages whose granular textual history remains unresolved.

The most accurate general finding is:

ChatGPT made substantial generative and expressive contributions to the manuscript, while the intellectual provenance, developmental provenance, and editorial disposition of the resulting material often involved distinct and sometimes iterative human–AI contributions.

The dossier therefore recognizes AI-generated expression directly while refusing to infer from sentence-level generation either the complete intellectual provenance of the passage or the legal authorship of the resulting work.

13. Direction, Selection, Revision, and Editorial Control

The production record supports continuing human direction and final human editorial control over Why Trust Oreskes’s Plumber?.

That finding must be stated precisely.

Human direction and editorial control are relevant to the developmental provenance of the work, but they are not treated as automatic substitutes for authorship of expression generated by ChatGPT.

The record therefore distinguishes among:

• direction — consequential human judgment concerning objectives, boundaries, sources, arguments, architecture, and permissible conclusions;
• selection — decisions concerning which proposed or generated material should be retained;
• revision — substantive modification of existing material;
• redirection — changes to the trajectory, function, conclusion, or placement of developing material;
• integration — incorporation of material into the larger manuscript;
• and editorial control — authority over the state ultimately authorized for publication.

These functions sometimes overlap, but they answer different provenance questions.

13.1 Continuing Direction

Human involvement continued well beyond the initial conception of the book.

The surviving record identifies repeated decisions by Capitanu concerning:

• what the book was trying to establish;
• what it was not permitted to claim;
• which controversies could enter at particular stages;
• which distinctions had to remain intact;
• what additional evidence was required;
• how opposing positions were to be treated;
• how individual chapters related to the larger argument;
• and whether particular proposed structures or conclusions were acceptable.

This continuing direction is significant because the project did not move from one originating prompt into autonomous AI production.

Human judgment remained active throughout the developmental chain.

13.2 Direction Before Drafting

Some of the strongest evidence of direction appears before prose generation.

The pre-drafting system required:

• definition of chapter function;
• research organization;
• identification of conceptual distinctions;
• argumentative constraints;
• source requirements;
• continuity relationships;
• risks;
• and validation before drafting.

At this stage, Capitanu could establish boundaries governing what later prose was permitted to do.

Examples include:

• distinguishing methodological from ontological naturalism;
• delaying premature treatment of Intelligent Design;
• requiring fair treatment of opposing positions;
• using Intelligent Design as an admissibility stress test rather than a predetermined verdict;
• preserving the distinction between admissibility and epistemic victory;
• and requiring a suspended conclusion where the evidence did not warrant more.

Such decisions are evidentially stronger than mere post hoc approval because they constrain development before the resulting prose exists.

13.3 Direction During Development

Human direction continued while ChatGPT was actively generating analysis and prose.

Capitanu could intervene when:

• an argument became too strong;
• a distinction was blurred;
• a treatment became one-sided;
• a proposed section occupied the wrong place;
• the evidentiary burden had been lowered;
• the relationship among chapters had become distorted;
• or a conceptual development departed from the governing purpose of the book.

Direction therefore functioned recursively.

It was not exhausted by the instructions supplied before generation.

13.4 Selection Is Relevant but Not Sufficient

The author selected material for inclusion in the manuscript.

That fact is relevant to the production history.

It is not treated as sufficient evidence that he authored every selected expression.

An editor, curator, commissioner, or other decision-maker may select among expressive alternatives without thereby becoming the originator of each selected sentence.

The same principle applies here.

Accordingly:

selection establishes editorial disposition; it does not automatically establish textual origin.

Its evidentiary significance increases where selection reflects substantive intellectual judgment, such as choosing between competing argumentative structures, rejecting an overstatement, preserving a conceptual distinction, or selecting a formulation because it satisfies a previously established intellectual requirement.

13.5 Rejection as Evidence of Governing Criteria

Rejection is particularly revealing because it demonstrates that generated output was not self-authorizing.

Capitanu could reject:

• a proposed structure;
• a conclusion;
• a framing;
• a placement;
• a passage;
• or an insufficiently qualified argument.

Where the reason for rejection is preserved, the event may reveal the intellectual standards governing the project.

For example, preserving a suspended verdict required rejecting routes that would have converted either methodological exclusion or criticism of exclusion into automatic substantive victory.

Rejection therefore constitutes more than absence of acceptance.

It may document active intellectual control.

13.6 Redirection

Redirection occurred when developing material was not simply discarded but given a different function.

This could involve:

• relocating a section;
• narrowing its scope;
• changing its role within a chapter;
• shifting an argument to a later chapter;
• preserving material while changing its conclusion;
• or replacing a proposed architecture with a new one.

One representative example concerns “inspectable trust.”

ChatGPT developed the material and proposed placement in Chapter 17.

Capitanu redirected it to Chapter 15 so that Chapter 17 could remain the concluding synthesis.

This was not a textual correction.

It was an architectural decision affecting the narrative and argumentative trajectory of the book.

13.7 Revision

Revision occurred throughout development.

It included:

• correction;
• qualification;
• expansion;
• compression;
• restructuring;
• replacement;
• and redrafting.

Some revisions were performed directly by Capitanu.

Others were performed by ChatGPT in response to human criticism or revised instructions.

Still others were iterative.

For provenance purposes, the important distinction is between:

who determined that revision was necessary

and

who generated the revised wording.

Those functions may belong to different participants.

13.8 Substantive Versus Mechanical Revision

Not all revision has equal evidentiary significance.

Mechanical correction includes matters such as:

• spelling;
• punctuation;
• formatting;
• minor stylistic cleanup.

Substantive revision alters:

• meaning;
• reasoning;
• evidentiary warrant;
• argumentative balance;
• architecture;
• or conceptual precision.

The record attributes greater provenance significance to the latter.

The human role documented in this project repeatedly extended into substantive revision and correction rather than remaining confined to mechanical cleanup.

13.9 Structural Control

The record also supports human control at the architectural level.

Capitanu determined or materially influenced:

• when controversial material entered the book;
• the relationship among Chapters 11 and 12;
• the placement of “inspectable trust”;
• the function of Chapter 15;
• the preservation of Chapter 17 as synthesis;
• and the later architecture of Chapters 13–17 as a five-movement sequence.

He also superseded proposed analytical arrangements where they did not fit the intended dramatic or philosophical structure.

These are consequential structural decisions.

13.10 Evidentiary Control

Human editorial control also operated upon the strength of claims.

Capitanu repeatedly required:

• additional nuance;
• stronger evidentiary support;
• preservation of uncertainty;
• differentiation between procedural and substantive conclusions;
• avoidance of premature endorsement;
• and fairer representation of opposing positions.

This form of control matters because it determines not merely what prose survives, but what the book is permitted to assert.

13.11 Canonicalization

By late August and early September, development increasingly shifted from conceptual expansion to canonical production control.

The record shows that corrected material existing somewhere in the working environment was not treated as equivalent to its presence in the authorized master.

Differences among:

• working documents;
• corrected drafts;
• production files;
• and canonical masters

had to be identified and reconciled.

The existence of a correction therefore did not matter operationally until it had been incorporated into the correct production version.

This stage provides strong evidence of final editorial and version-control authority.

13.12 Format Separation

The work subsequently existed in separately controlled production forms, including:

• print;
• EPUB;
• and audiobook.

Each required its own production review.

The separation of formats is relevant because final authorization was not a single abstract event.

Individual production manifestations had to be checked and accepted.

The fact that a text had reached mature manuscript form did not eliminate later editorial control.

13.13 Audiobook as Ancillary Evidence

Audiobook production provides ancillary evidence of the same control pattern.

Generated audio was subjected to:

• pronunciation testing;
• listening;
• correction;
• rejection of defective output;
• regeneration;
• and explicit approval.

This does not establish authorship of the written manuscript.

It does corroborate a broader project-level practice:

AI-generated output was not treated as self-authorizing merely because it had been successfully generated.

Human acceptance remained a necessary production event.

13.14 Responsibility for the Published Work

Capitanu ultimately accepted responsibility for the published work and exercised the authority to determine:

• what entered the manuscript;
• what required further work;
• what was removed;
• what was moved;
• which version was canonical;
• and which production state was authorized.

This responsibility is relevant to the descriptive classification adopted by this dossier.

It is not treated as retroactive proof that all AI-generated expression became human-generated expression.

13.15 Editorial Control and Textual Provenance

The governing distinction is:

editorial control does not equal textual provenance.

A human may exercise final authority over a passage while the wording remains AI-generated.

Conversely, a human-originating concept may appear in an AI-generated formulation.

The record therefore preserves both dimensions separately.

13.16 Editorial Control and Authorship

A second distinction is equally important:

control over publication is not equivalent to authorship.

Final publication authority alone would be insufficient to establish substantive human intellectual contribution.

The finding in this project is stronger because editorial control forms only one part of a broader documented pattern that also includes:

• intellectual origin;
• continuing direction;
• substantive conceptual intervention;
• research direction;
• argumentative constraints;
• structural decisions;
• rejection;
• redirection;
• and revision.

Editorial control is therefore corroborative within a larger developmental record.

13.17 Direction and Control Finding

The surviving evidence supports the conclusion that human control was substantive and continuous.

Capitanu determined:

• when drafting could proceed;
• whether arguments required correction;
• whether structures were retained or replaced;
• whether conclusions exceeded their evidentiary warrant;
• whether material should be accepted, rejected, redirected, revised, or relocated;
• and when chapters and production versions could be treated as canonical.

ChatGPT exercised substantial influence upon the material available for those decisions and generated substantial material ultimately retained.

It did not possess independent authority to determine the final state of the manuscript or its publication.

The most accurate conclusion is therefore:

the work developed through substantial AI contribution within a process of continuing human intellectual direction and final human editorial control.

This finding concerns the organization and disposition of the work.

It does not erase the distinct provenance of AI-generated expression.

14. Chronology

The following chronology summarizes the principal provenance-relevant stages in the development of Why Trust Oreskes’s Plumber?.

Its purpose is not to reproduce every developmental conversation or editorial event.

It identifies the principal points at which the surviving record permits meaningful conclusions concerning:

• intellectual origin;
• architectural development;
• research and pre-drafting organization;
• human direction;
• AI contribution;
• iterative development;
• and final editorial control.

Later manuscript or production evidence may corroborate the persistence of earlier concepts, but it is not used by itself as proof of their original provenance.

14.1 26–27 May 2026 — Formation of the Project

Evidentiary status: PARTIALLY DOCUMENTED / EARLIEST RECONSTRUCTED STAGE

The project developed around Naomi Oreskes’s treatment of scientific trust and the plumber analogy.

By this stage, Capitanu had established the central critical orientation:

the case for informed trust in science should be taken seriously, while the mechanisms invoked to warrant that trust should themselves remain open to inspection.

The emerging field included:

• philosophical assumptions;
• consensus;
• peer review;
• conflicts of interest;
• historical failures;
• expertise;
• and public trust.

The rhetorical posture was established as restrained or sober satire rather than anti-scientific polemic.

The plumber motif was already part of the developing project.

14.2 27 May 2026 — Initial Seventeen-Chapter Architecture

Evidentiary status: DOCUMENTED IN THE RECONSTRUCTED DEVELOPMENTAL RECORD

ChatGPT proposed an organized seventeen-chapter structure transforming the initial intellectual field into a book-length progression.

This constitutes a substantial AI contribution to book-level architecture.

Capitanu accepted the seventeen-chapter framework and subsequently directed its development and modification.

The provenance distinction established at this point is:

human intellectual initiation

and

substantial AI macro-architectural contribution.

14.3 27 May 2026 — Early Drafting Constraints

Evidentiary status: DOCUMENTED

Capitanu restricted the initial drafting stage to the Prologue and Chapters 1–2.

He directed that Intelligent Design and anti-Intelligent-Design sources not be introduced prematurely.

Oreskes was to receive fair treatment rather than be constructed merely as a rhetorical target.

ChatGPT subsequently generated substantial prose within those constraints.

This establishes an early pattern of:

human intellectual/editorial direction → AI textual execution and development.

14.4 27 May 2026 — Establishment of the Pre-Drafting Protocol

Evidentiary status: DOCUMENTED

A recurring working protocol was established under which drafting would be preceded by consultation of project documentation and identification of:

• function;
• thesis;
• internal structure;
• relevant dossiers;
• source requirements;
• footnote claims;
• bibliography;
• and sensitive argumentative issues.

Drafting was to begin only after validation.

This event is important because it shows that the production method was structured from the beginning rather than retrospectively reconstructed.

14.5 Late May 2026 — Early Prose Generation

Evidentiary status: DOCUMENTED IN THE RECONSTRUCTED RECORD

ChatGPT participated directly in substantial drafting for the Prologue and early chapters.

Its contribution included development of material concerning:

• the plumber analogy;
• informed trust;
• consensus;
• scientific community;
• philosophical diversity;
• and the requirement that warranted trust remain inspectable.

The AI role had already exceeded proofreading or conventional editing.

14.6 3 June 2026 — Part II Pre-Drafting: Chapters 3–4

Evidentiary status: DOCUMENTED

The project used updated continuity documentation and pre-drafting analysis involving:

• page targets;
• architecture;
• constraints;
• dossiers;
• and provenance-sensitive planning.

Capitanu required preservation of the distinction between methodological and ontological naturalism and deliberately limited the early introduction of Intelligent Design.

14.7 14 June 2026 — Part III Pre-Drafting: Chapters 5–6

Evidentiary status: DOCUMENTED

The preparatory stage included multiple working documents addressing:

• page allocation;
• continuity;
• sources;
• footnotes;
• bibliography;
• risks;
• transitions;
• and internal architecture.

This confirms that pre-drafting organization was becoming a recurring production method.

14.8 21 June 2026 — Part IV Pre-Drafting: Chapters 7–8

Evidentiary status: DOCUMENTED

Preparatory work included:

• target pages;
• citations;
• continuity;
• and architecture before drafting.

Capitanu introduced the requirement to treat manufactured doubt while also extending conflicts-of-interest analysis beyond financial interests to include:

• prestige;
• reputation;
• career incentives;
• and institutional authority.

14.9 24 June 2026 — Part V Pre-Drafting: Chapters 9–10

Evidentiary status: DOCUMENTED

The working stage again included:

• multiple documents;
• page targets;
• dossiers;
• sources;
• architecture;
• footnote needs;
• continuity;
• and risks.

The record explicitly distinguished human planning inputs from ChatGPT proposals.

Drafting remained subject to prior validation.

14.10 24 June 2026 — Part VI Pre-Drafting: Chapters 11–12

Evidentiary status: DOCUMENTED

The working environment included:

• Continuity Brief;
• Analytical Dossiers;
• Oreskes source material;
• existing Manuscript Draft.

The preparatory stage addressed:

• function;
• dossiers;
• constraints;
• section architecture;
• footnotes;
• bibliography;
• risks;
• and page strategy.

The checkpoint explicitly preceded substantial prose generation.

This stage later became especially important for the representative reconstruction of Chapter 11.

14.11 Late June 2026 — Intelligent Design as an Admissibility Stress Test

Evidentiary status: DOCUMENTED / STRONGLY CORROBORATED

Capitanu established a governing constraint that Intelligent Design would function as a stress test of scientific admissibility rather than as a predetermined conclusion.

The discussion was required to preserve the distinction between:

procedural admissibility

and

epistemic success.

The book was to avoid both automatic anti-ID and automatic pro-ID conclusions.

14.12 30 June 2026 — “The Right to Be Wrong”

Evidentiary status: DOCUMENTED HUMAN CONCEPTUAL INTERVENTION

Capitanu introduced the principle that science possesses a conditional right to be wrong insofar as error remains exposed to correction.

ChatGPT subsequently developed and formulated the principle within Chapter 11.

The provenance distinction is:

governing concept — HUMAN

analytical development — AI / ITERATIVE

expressive formulation — AI.

14.13 3–4 July 2026 — Expansion of Chapter 11

Evidentiary status: DOCUMENTED ITERATIVE DEVELOPMENT

The developing architecture included:

• “The Right to Be Wrong”;
• examination before dismissal;
• procedural admissibility versus epistemic victory;
• minimum seriousness;
• fallible scientific borders;
• refutation versus expulsion by classification;
• curricular distinctions;
• fertile error;
• and satirical plumber material.

Capitanu continued to require the suspended verdict and avoidance of pro-ID apologetics.

14.14 6 July 2026 — Part VII Pre-Drafting: Chapters 13–15

Evidentiary status: DOCUMENTED

The preparatory environment included:

• document use;
• page targets;
• function and thesis;
• structure;
• dossiers;
• continuity;
• footnotes;
• bibliography;
• risks;
• and page strategy.

Chapter 13 received additional planning involving:

• argument mapping;
• documentary evidence;
• dramaturgy;
• analogies;
• and section/page budgeting.

14.15 14–15 July 2026 — Part VIII Pre-Drafting: Chapters 16–17

Evidentiary status: DOCUMENTED

Planning addressed:

• thesis;
• section architecture;
• dossiers;
• continuity constraints;
• footnote claims;
• bibliography;
• risks;
• page allocation;
• and search checks.

Capitanu superseded an initial analytical arrangement and imposed a broader architecture in which Chapters 13–17 functioned as a five-movement unit, Chapter 15 as climax, and Part VIII as philosophical and civic resolution.

14.16 16 July 2026 — Philosophical Charity and Inspectable Trust

Evidentiary status: DOCUMENTED IN RECONSTRUCTED CONVERSATIONAL RECORD; LATER PERSISTENCE CORROBORATED

Capitanu raised the question whether science requires philosophical charity or inspection.

ChatGPT developed the intervention into a proposed section and further developed the concept of inspectable trust.

ChatGPT proposed placement in Chapter 17.

Capitanu redirected it to Chapter 15.

The sequence therefore separates:

• human initiating question;
• AI/iterative conceptual development;
• AI placement proposal;
• human final architectural decision.

14.17 21 July 2026 — Manuscript Checkpoint for the Suspended Verdict

Evidentiary status: DOCUMENTED BY SURVIVING MANUSCRIPT

A surviving manuscript checkpoint shows that the distinction between admissibility and epistemic victory had become embedded in the work.

Neither methodological naturalism nor Intelligent Design received automatic epistemic privilege merely by procedural rule.

This provides evidence independent of later recollection.

14.18 2 August 2026 — From “Meteors” to Research Program

Evidentiary status: DOCUMENTED IN RECONSTRUCTED CONVERSATIONAL RECORD; LATER MANUSCRIPT CORROBORATED

Capitanu introduced the “meteors” analogy for isolated Intelligent Design arguments that do not yet constitute a mature research program.

He asked what would be required for design to become scientifically competitive.

ChatGPT developed this into a framework involving:

• discriminating predictions;
• falsifiability;
• cumulative research performance;
• comparative explanatory performance;
• and avoidance of mere reliance upon explanatory gaps.

This is a representative case of:

human conceptual initiation → substantial AI analytical development → manuscript integration.

14.19 8–10 August 2026 — Public Description of the Emerging Architecture

Evidentiary status: CORROBORATIVE

Materials prepared for public or production use described the book as:

• a critical response to Why Trust Science?;
• organized around the plumber analogy;
• written as restrained or sober satire;
• using Intelligent Design as a stress test rather than required conclusion;
• and developing inspectable trust as a central theme.

Because these materials postdate the underlying development, they corroborate persistence but are not used to establish first provenance.

14.20 14 August 2026 — Mature Seventeen-Chapter Production Structure

Evidentiary status: DOCUMENTED

Production records show a mature seventeen-chapter structure.

The broad architecture proposed in May had survived.

Individual chapters, sections, and functions had nevertheless undergone extensive development.

This establishes structural continuity, not exclusive provenance of the content within each chapter.

14.21 20 August 2026 — Institutional Mediation and Future Technologies

Evidentiary status: DOCUMENTED

Capitanu introduced COVID-19 as a case illustrating the path through which scientific findings become consequential for citizens through:

• expert bodies;
• private institutions;
• governmental decision-making;
• and public authority.

He extended the concern to future technologies including:

• artificial intelligence;
• biotechnology;
• and biohacking.

ChatGPT systematized the concern into a more explicit chain involving:

evidence → interpretation → institutional judgment → private interests → policy → public authority.

This broadened the book from epistemic trust toward civic consequences.

14.22 Late August 2026 — Consolidation of the Mature Manuscript

Evidentiary status: DOCUMENTED

By late August, the seventeen-chapter structure had become mature.

Development increasingly shifted from expansion of intellectual architecture toward:

• manuscript reconciliation;
• production preparation;
• version control;
• and canonicalization.

14.23 28 August–3 September 2026 — Canonicalization and Version Control

Evidentiary status: DOCUMENTED

Differences among:

• working documents;
• revised materials;
• and production masters

required identification and correction.

The presence of corrected material in a working file was not treated as equivalent to its presence in the authorized master.

This stage provides evidence of substantive editorial and production control.

14.24 September 2026 — Publication Masters and Format Separation

Evidentiary status: DOCUMENTED

The work entered separately controlled production states for:

• print;
• EPUB;
• and audiobook.

Production files identified Daniel Capitanu as author and Les Éditions du Rasoir as publisher.

These records establish how the completed work was presented and controlled at publication stage.

They do not independently establish the intellectual or textual provenance of individual components.

14.25 September 2026 — Audiobook Production

Evidentiary status: DOCUMENTED; ANCILLARY TO TEXTUAL PROVENANCE

Audiobook production added another layer of human review.

Generated audio was:

• tested;
• listened to;
• corrected;
• rejected where defective;
• regenerated;
• and explicitly approved.

This does not establish authorship of the underlying written text.

It does corroborate the broader production pattern that generated AI output was subject to human quality control and authorization.

14.26 Chronological Finding

Across the period from late May through September 2026, the surviving record supports a developmental sequence characterized by:

human intellectual initiation

→

substantial AI macro-architecture

→

systematic human–AI research and pre-drafting development

→

substantial AI analytical and textual generation

→

continuing human conceptual direction, rejection, correction, and structural redirection

→

iterative development

→

human canonicalization and final editorial control.

The chronology does not imply that every chapter followed an identical production path.

Its significance lies in demonstrating that the documented patterns identified elsewhere in this dossier occurred across the development of the work rather than appearing only in isolated favorable examples.

15. Representative Reconstruction: Chapter 11, “The Forbidden Inference”

Chapter 11, “The Forbidden Inference,” is selected as the principal representative provenance reconstruction because it occupies a conceptually central position in Why Trust Oreskes’s Plumber?.

It opens Part VI, where Intelligent Design functions as a stress test of the book’s broader inquiry into:

• scientific admissibility;
• methodological rules;
• evidentiary evaluation;
• classification;
• refutation;
• fallibility;
• and warranted trust.

The chapter is therefore particularly well suited to testing whether the developmental record supports a simple model of:

human topic → AI-generated chapter

or instead reveals a more complex production chain involving prior intellectual organization, AI development, continuing human intervention, and mixed textual provenance.

The reconstruction is deliberately representative rather than sentence-by-sentence.

Its purpose is to determine whether the surviving evidence permits a reliable account of the chapter’s production process and whether that account materially supports or contradicts the broader provenance findings of this dossier.

15.1 Analytical Stages

The reconstruction examines six stages:

INPUT HUMAN

→

INPUT AI

→

VALIDATED PLAN

→

PROSE GENERATED

→

HUMAN INTERVENTION

→

CANONICAL RESULT

These categories describe stages of production.

They do not constitute presumptions concerning legal authorship.

15.2 Earlier Intellectual Antecedents

The intellectual problem of Chapter 11 predates its mature prose.

Before substantial drafting, the project had already developed several distinctions relevant to the chapter.

Among them were:

• methodological naturalism versus ontological naturalism;
• classification versus substantive refutation;
• admissibility versus epistemic success;
• and Intelligent Design as a stress test rather than a predetermined conclusion.

Earlier preparatory work had also treated controversial cases within evolutionary debate as potential stress tests of methodological boundaries rather than as automatic evidence for Intelligent Design.

The chapter therefore emerged from an intellectual environment that already contained a developed problem concerning the relationship between:

methodological rule

and

evidentiary judgment.

15.3 Standing Production Protocol

Before Chapter 11 entered mature development, the project was operating under an explicit pre-drafting protocol.

Four principal working resources governed the process:

1. Continuity Brief — project-level framework and continuity control;
2. Analytical Dossiers — research, analytical development, and planning;
3. Oreskes source material — primary-source reference;
4. Manuscript Draft So Far — continuity, repetition, and integration control.

The production procedure required consideration of:

• sources;
• risks;
• architecture;
• page strategy;
• footnotes;
• bibliography;
• and continuity

before drafting proceeded.

Drafting was subject to validation rather than commencing automatically.

This protocol predates Chapter 11 and therefore constitutes evidence about the production environment within which the chapter was later developed.

Its provenance is best classified as:

ITERATIVE SYSTEM WITH DOCUMENTED HUMAN ADOPTION AND CONTROL.

15.4 Earlier Research Reservoir

By mid-June, the Analytical Dossiers included or contemplated a research reservoir concerning the genealogy and philosophical status of methodological naturalism.

The source field included historical and philosophical figures and debates relevant to:

• scientific demarcation;
• naturalism;
• evolutionary controversy;
• legal and institutional boundaries;
• and the philosophical status of methodological rules.

The purpose was not merely to assemble authorities favoring a predetermined conclusion.

The relevant research problem was whether a rule governing scientific explanation should itself remain historically and philosophically inspectable.

This research architecture existed before the mature chapter prose.

The precise provenance of every individual source selection remains only partly reconstructable.

Accordingly:

pre-existing research architecture — DOCUMENTED / CORROBORATED

granular source-selection provenance — PARTLY UNRESOLVED.

15.5 Methodological Versus Ontological Naturalism

Before Chapter 11 was drafted, the project had already established that methodological and ontological naturalism were not to be treated as interchangeable.

The distinction was consequential.

Methodological naturalism could be analyzed as a rule defining the permissible form of scientific explanation without automatically being treated as proof of the ontological proposition that only natural causes exist.

Capitanu is documented as requiring preservation of this distinction.

ChatGPT participated substantially in its later analysis, organization, and expression.

The most accurate classification is therefore:

HUMAN CONSTRAINT + AI DEVELOPMENT → ITERATIVE IMPLEMENTATION.

15.6 Classification Versus Refutation

A second pre-existing problem concerned the difference between excluding a claim by classification and rejecting it through substantive argument.

The project had already begun to resist two conflations:

methodologically inadmissible ≠ demonstrated false

and

permitted to be examined ≠ demonstrated true.

This distinction became one of the central epistemic structures of Chapter 11.

The pre-drafting existence of the distinction is corroborated.

Its exact first formulation remains less certain.

15.7 Intelligent Design as Stress Test

By the time Chapter 11 entered active development, Intelligent Design had acquired a deliberately constrained function.

It was not to be treated as a conclusion the book was required to prove.

Instead, it served as a stress test of the broader epistemological framework.

The governing question became:

Can the rules by which a controversial inference is excluded themselves be inspected without converting that inspection into endorsement of the inference?

This required preserving two different questions:

1. Is the inference entitled to examination?
2. Does the inference survive examination?

The second question was reserved substantially for Chapter 12.

The provenance finding is:

HUMAN GOVERNING CONSTRAINT — STRONGLY SUPPORTED

with

SUBSTANTIAL AI PARTICIPATION IN DEVELOPMENT.

15.8 The 24 June 2026 Pre-Drafting Checkpoint

On 24 June, active development turned explicitly to Part VI, Chapters 11–12.

The working record identifies the four governing resources:

• Continuity Brief;
• Analytical Dossiers;
• Oreskes source material;
• existing Manuscript Draft.

Part VI was provisionally allocated approximately 42 body-text pages:

• Chapter 11 — approximately 18 pages;
• Chapter 12 — approximately 24 pages.

For Chapter 11, the planning architecture included:

• Oreskes;
• methodological naturalism;
• distinctions relevant to Intelligent Design;
• scientific demarcation;
• Kitzmiller and institutional boundary questions;
• classification versus refutation;
• and transition into the more substantive examination reserved for Chapter 12.

This is direct evidence that the chapter possessed an articulated research and argumentative architecture before mature prose generation.

The production therefore cannot accurately be represented as:

“Write a chapter about Intelligent Design” → generated chapter.

An intervening planning layer is documented.

15.9 Provenance of the 24 June Architecture

The existence of the architecture is more securely established than exclusive authorship of every component.

The record supports the following distinctions.

HUMAN

Capitanu required:

• continued use of the documentary system;
• continuity control;
• attention to provenance;
• source discipline;
• page strategy;
• footnote and bibliography planning;
• risk assessment;
• and validation before drafting.

He had also imposed substantive constraints concerning:

• methodological versus ontological naturalism;
• avoidance of premature Intelligent Design apologetics;
• and preservation of the broader epistemic framework.

AI

ChatGPT substantially participated in converting accumulated materials and constraints into:

• section architecture;
• analytical sequence;
• source deployment;
• transition strategy;
• and a workable chapter plan.

ITERATIVE

The resulting architecture cannot accurately be reduced to either:

human plan merely transcribed by AI

or

AI plan merely approved by the human participant.

It emerged through interaction among pre-existing human constraints, accumulated collaborative documentation, and substantial AI analytical organization.

UNRESOLVED

The exact first origin of every subordinate component of the 24 June plan has not been reconstructed.

15.10 The Suspended Verdict

On 29 June, the epistemic status of the Intelligent Design discussion became explicit.

The governing verdict was to remain suspended.

The book was to be neither pro-ID nor anti-ID in its controlling conclusion.

The relevant question was whether exclusion was:

• argued;
• explicit;
• symmetrical;
• and inspectable.

This decision excluded two simpler routes:

methodological exclusion → therefore Intelligent Design is false

and

criticism of methodological exclusion → therefore Intelligent Design is true.

The chapter was required to preserve the space between admissibility and epistemic success.

Classification:

HUMAN — DOCUMENTED.

15.11 The “Right to Be Wrong”

On 30 June, Capitanu introduced the further principle that science possesses a conditional right to be wrong, provided that error remains exposed to correction.

This intervention added another dimension to the admissibility problem.

The issue became not only whether methodological borders can be inspected, but whether a scientific culture committed to fallibility can justify preventing a sufficiently serious claim from entering the arena in which error becomes visible and correctable.

ChatGPT subsequently developed and formulated this principle.

Because this episode permits unusually clear provenance separation, it receives its own detailed treatment in Section 16.

15.12 Expansion of the Architecture: 3–4 July

By early July, the chapter architecture had expanded to include:

• “The Right to Be Wrong”;
• “Not Belief, but Examination”;
• procedural admissibility versus epistemic victory;
• minimum seriousness;
• fallible scientific borders;
• refutation rather than expulsion;
• curricular distinctions;
• fertile error;
• and restrained plumber satire.

Capitanu simultaneously required preservation of the suspended verdict and avoidance of pro-ID apologetics.

The developmental status of this material is important.

The record does not establish that every element of the 3–4 July architecture existed strictly before the first prose generation.

Some may have developed during drafting and revision.

The appropriate classification is therefore:

ITERATIVE DEVELOPMENT

with

exact temporal relationship to earliest prose partly unresolved.

15.13 Before-Generation Finding

At or before the threshold of substantial prose development, the project contained:

• an established production protocol;
• a continuity system;
• a research reservoir;
• the methodological/ontological naturalism distinction;
• classification versus refutation;
• Intelligent Design as an admissibility stress test;
• page and structural planning;
• the Chapter 11–12 division of labor;
• a suspended verdict;
• and the human-originating “right to be wrong” principle.

This establishes more than a human-selected topic.

At the same time, the pre-drafting architecture was not wholly human-created.

ChatGPT made substantial analytical and organizational contributions.

The best-supported reconstruction before mature generation is therefore:

human-originating intellectual constraints and interventions

accumulated collaborative research documentation

substantial AI analytical organization

human validation and consequential redirection

→

chapter-level architecture from which mature prose developed.

15.14 What ChatGPT Received at Generation Stage

By the time substantial Chapter 11 prose development was underway, ChatGPT was not responding to an isolated general request.

The working environment already contained:

• the book-level problem of warranted and inspectable trust;
• manuscript continuity;
• methodological versus ontological naturalism;
• accumulated research on naturalism and demarcation;
• classification versus substantive refutation;
• Intelligent Design as stress test;
• admissibility versus epistemic victory;
• the planned division between Chapters 11 and 12;
• the suspended verdict;
• and the developing principle of scientific fallibility and correction.

AI contribution must therefore be assessed against this substantial pre-existing intellectual environment.

15.15 Substantial AI Development During Generation

Within that environment, ChatGPT made substantial analytical contributions.

Representative developments included:

• examination before dismissal;
• a threshold of minimum seriousness;
• scientific boundaries as fallible rather than automatically self-validating;
• refutation versus expulsion by classification;
• procedural admissibility versus epistemic victory;
• error as potentially productive when exposed to correction;
• curricular distinctions;
• symmetry as a standard of inspection;
• and satirical formulations returning the epistemic problem to the plumber metaphor.

These were not merely mechanical expansions of fully specified sentences.

They constitute:

SUBSTANTIAL AI ANALYTICAL DEVELOPMENT.

15.16 Substantial AI Drafting

The surviving record also establishes substantial ChatGPT participation in drafting.

The prose developed the chapter around the proposition that classification is insufficient.

It examined two opposing dangers.

On one side:

incomplete naturalistic explanation must not become an automatic invitation to insert design.

On the other:

methodological naturalism must not become an authorized refusal to examine certain causal questions merely by classification.

The chapter therefore separated:

permission to examine an inference

from

reason to accept that inference.

This distinction became a governing feature of the canonical chapter.

The record does not presently justify attributing the complete chapter prose exclusively to either Capitanu or ChatGPT.

Passage-level provenance is mixed and only partly reconstructed.

15.17 Continuing Human Intervention

Human intervention continued during and after substantial drafting.

Capitanu:

• introduced the “right to be wrong” principle;
• maintained the suspended verdict;
• required that inspection not collapse into endorsement;
• preserved the Chapter 11–12 division of labor;
• and continued to constrain the chapter against pro-ID apologetics.

By early July, Chapter 11 had become the conceptual nucleus for the more technical examination that followed in Chapter 12.

What remains less precisely established is:

• the quantity of sentence-level rewriting directly performed by Capitanu;
• the full inventory of rejected AI-generated passages;
• and the degree to which first-generation wording changed before canonicalization.

Those matters remain partly unresolved.

15.18 The Canonical Result

Surviving July manuscript versions contain Chapter 11 in substantially developed form.

The canonical argument asks whether naturalism functions as:

• a practical methodological rule;
• a philosophical decree;
• or some combination of the two.

The chapter rejects two symmetrical shortcuts.

It does not allow difficulty in a naturalistic account automatically to establish design.

It also does not allow methodological classification of design as inadmissible to substitute automatically for evidentiary refutation.

The resulting chapter preserves the distinction between:

procedural admissibility

and

epistemic victory.

Substantive examination is deliberately carried forward into Chapter 12.

The September canonical EPUB retains Chapter 11 as “The Forbidden Inference” at the opening of Part VI, Intelligent Design as a Stress Test.

15.19 Provenance Matrix

The representative reconstruction supports the following findings.

Book-level trust problem Human origin supported; AI developmental contribution substantial.

Methodological versus ontological naturalism Human constraint documented; AI development substantial.

Intelligent Design as admissibility stress test Human governing direction strongly supported; AI analytical development substantial.

Suspended verdict Human governing constraint documented.

“Right to be wrong” principle Human conceptual origin documented; AI development and formulation substantial.

Pre-drafting architecture Collaborative / iterative; human constraints and validation combined with substantial AI organization.

Chapter prose Substantial AI drafting; mixed textual provenance; complete sentence-level reconstruction not claimed.

Final relationship between Chapters 11 and 12 Human editorial control supported; AI structural participation also documented.

15.20 Adversarial Finding

Chapter 11 does not support the description:

Capitanu supplied a topic and ChatGPT independently devised and wrote the intellectual argument.

The evidence shows substantive human intellectual inputs before and during development.

Nor does it support:

Capitanu wrote the substantive chapter and ChatGPT merely polished it.

ChatGPT contributed substantially to:

• architecture;
• conceptual development;
• formulation;
• analytical expansion;
• and prose generation.

The best-supported reconstruction is:

human intellectual problem and governing constraints

→

collaborative pre-drafting development

→

substantial AI analytical and textual generation

→

continuing human conceptual intervention and editorial control

→

collaboratively developed canonical chapter.

15.21 Significance of the Representative Reconstruction

The Chapter 11 audit provides concrete evidence that the broader provenance account is not based only on project-level recollection or final manuscript ownership.

It shows identifiable human intellectual contribution before and during AI generation while simultaneously documenting substantial AI analytical and expressive contribution.

The chapter therefore supports the broader conclusion that the production process cannot accurately be reduced either to autonomous AI generation from high-level human prompts or to conventional editing of independently human-authored prose.

At the same time, this is a representative audit.

It is not a claim to exhaustive forensic reconstruction of every sentence in Chapter 11 or every chapter in the book.

Its purpose is reasonable evidentiary sufficiency: to demonstrate the nature of the production process with enough granularity that the broader account can be independently tested.

16. The “Right to Be Wrong” Case

The “Right to Be Wrong” provides one of the clearest individual provenance cases in the entire reconstruction.

It is valuable because the surviving developmental record permits three analytically distinct questions to be answered separately:

1. Who introduced the governing concept?
2. Who substantially developed the concept into an argument?
3. Who generated the representative expressive formulation?

The answers are not identical.

This makes the episode a useful demonstration of why conceptual provenance, developmental provenance, and textual provenance cannot be collapsed into one attribution.

16.1 The Conceptual Intervention

On 30 June 2026, Daniel Capitanu introduced the principle that science possesses a conditional right to be wrong, provided that its errors remain visible to correction.

The principle was connected explicitly to controversial inferences, including Intelligent Design.

The intervention did more than supply a phrase or topic.

It established a rule governing the epistemic problem.

Scientific fallibility is productive only where error remains open to:

• inspection;
• criticism;
• correction;
• and possible refutation.

The resulting question was therefore:

Can a scientific culture committed to fallibility justify excluding a sufficiently serious claim before that claim has entered the arena in which its possible error can be exposed?

The human contribution is classified:

GOVERNING CONCEPT — HUMAN — DOCUMENTED.

16.2 Relation to the Existing Admissibility Problem

Before this intervention, Chapter 11 already contained a developing problem concerning:

• methodological boundaries;
• admissibility;
• classification;
• refutation;
• and the distinction between permission to examine and entitlement to believe.

The “right to be wrong” did not replace that architecture.

It deepened it.

The earlier question had been:

Can methodological exclusion substitute for evidentiary refutation?

The new principle added:

What happens to scientific fallibility if some serious claims are excluded from the arena of possible correction before they can be examined?

The intervention therefore connected:

fallibility

with

procedural openness to examination.

16.3 AI Analytical Development

ChatGPT subsequently developed the principle into a broader Chapter 11 argument.

Its contribution included connecting the human-originating concept with:

• examination before dismissal;
• minimum seriousness;
• fallible methodological borders;
• productive error;
• classification versus refutation;
• and the distinction between examination and belief.

This was not merely stylistic elaboration.

It constituted substantive analytical development.

Accordingly:

ANALYTICAL DEVELOPMENT — AI / ITERATIVE.

The term iterative remains appropriate because development occurred within the chapter architecture and governing constraints already established through the human–AI process.

16.4 Expressive Formulation

One resulting formulation was:

“The right to be wrong is not the right to be believed. It is the right to be examined before being dismissed.”

The developmental record attributes this expressive formulation to ChatGPT.

The formulation condenses several distinctions:

• fallibility does not create entitlement to acceptance;
• permission to examine is distinct from reason to believe;
• exclusion is distinct from refutation;
• and a claim may deserve examination without deserving epistemic victory.

The appropriate classification is:

EXPRESSIVE FORMULATION — AI.

16.5 Why the Attribution Must Be Split

Two opposite attribution errors are possible.

The first would be:

ChatGPT generated the mature formulation; therefore ChatGPT originated the entire concept.

That would erase the documented human introduction of the governing idea.

The second would be:

Capitanu originated the concept; therefore the resulting developed argument and formulation are wholly his intellectual and expressive contribution.

That would erase substantial AI analytical and textual development.

Both would distort the documentary history.

The more accurate provenance sequence is:

GOVERNING CONCEPT — HUMAN

→

ANALYTICAL DEVELOPMENT — AI / ITERATIVE

→

EXPRESSIVE FORMULATION — AI

→

HUMAN EVALUATION AND INTEGRATION WITHIN THE CHAPTER.

16.6 The Role of Human Constraint

The AI development did not occur in an unconstrained conceptual space.

The chapter was already subject to:

• the suspended verdict;
• the requirement that criticism of exclusion not establish Intelligent Design;
• the distinction between methodological and ontological naturalism;
• and the separation between procedural admissibility and epistemic success.

The “right to be wrong” therefore had to operate symmetrically.

It could not become a device for granting epistemic privilege to Intelligent Design.

Its function was procedural:

error must remain examinable before correction is possible.

This constraint shaped the meaning of the resulting AI-developed argument.

16.7 Minimum Seriousness

The AI development also introduced or expanded the idea that examination is not owed indiscriminately to every claim.

A controversial inference must meet some threshold of minimum seriousness before procedural examination becomes relevant.

This qualification is important.

Without it, the “right to be wrong” could collapse into an obligation to investigate any proposition merely because it has been asserted.

The developed argument therefore distinguished between:

• openness to serious examination;
• and indiscriminate epistemic permissiveness.

This is an example of substantive AI analytical extension within the governing framework.

16.8 Productive Error

Another development concerned the epistemic usefulness of error.

If an incorrect claim enters an environment where it can be:

• examined;
• tested;
• criticized;
• and corrected,

its failure may itself contribute to knowledge.

Error may therefore be epistemically productive.

But where a claim is excluded categorically before substantive examination, one loses the possibility of distinguishing:

• error exposed through examination;
• from exclusion imposed through prior classification.

This analytical development expanded the conceptual reach of the original human intervention.

16.9 Symmetrical Inspection

The principle also became linked to symmetrical standards.

If methodological naturalism is inspectable, design claims must also remain inspectable.

But symmetry does not mean equal evidentiary merit.

Naturalistic explanation is not automatically protected from scrutiny by institutional status.

Design inference is not automatically privileged because its exclusion has been questioned.

The proper symmetry concerns:

the standard of inspection

rather than

the outcome of inspection.

This was substantially developed by ChatGPT under the existing human constraint of suspended verdict.

16.10 Relation to the Plumber Metaphor

The principle was also translated into the established plumber language of the book.

ChatGPT proposed formulations linking refusal to inspect a leak with failure to establish that a plumber’s diagnosis was wrong.

This expressive development brought the chapter’s abstract epistemological problem back into the governing satire of the book.

The broader metaphor predated the particular formulation.

The specific phrasing was an AI contribution.

Again, conceptual framework and expression have different provenance.

16.11 Human Selection and Integration

AI-generated formulations did not enter the developing chapter automatically.

They were subject to the ongoing editorial process.

The “right to be wrong” material was retained because it satisfied the chapter’s governing purpose and constraints.

Human selection and integration are therefore part of the developmental history.

They do not retroactively make the AI-generated expression human-generated.

The appropriate description remains:

AI-generated expression selected and integrated within a human-directed argumentative framework.

16.12 Persistence Into the Chapter Architecture

By 3–4 July, “The Right to Be Wrong” had become an explicit component of the Chapter 11 architecture.

It was accompanied by related elements including:

• “Not Belief, but Examination”;
• procedural admissibility versus epistemic victory;
• minimum seriousness;
• fallible borders;
• fertile error;
• and symmetry.

This shows that the concept was not an isolated rhetorical flourish.

It became a structural element of the chapter’s argument.

16.13 Evidentiary Value of the Case

This episode is especially valuable because it prevents both exaggeration and minimization.

It prevents exaggeration of human contribution because the developed analysis and representative formulation were not wholly human-originating.

It prevents exaggeration of AI contribution because the governing concept did not originate with ChatGPT.

The case therefore demonstrates concretely why provenance must be reconstructed across separate layers.

16.14 Provenance Layers

The “Right to Be Wrong” can be represented as follows.

Conceptual provenance Human.

Initial governing function Human.

Analytical development Substantial AI / iterative.

Integration with admissibility and fallibility Iterative, with substantial AI development under human constraints.

Representative expressive formulation AI.

Retention and integration into Chapter 11 Human editorial selection within the iterative process.

Final canonical function Part of the collaboratively developed chapter architecture.

16.15 Why the Case Matters Beyond Chapter 11

The significance of the case extends beyond the individual concept.

It demonstrates a recurring structural feature of the book’s production.

A human-originating intellectual intervention could be transformed by ChatGPT into:

• a developed argument;
• a conceptual network;
• a rhetorical formulation;
• and publishable prose.

The author could then:

• evaluate;
• constrain;
• redirect;
• select;
• and integrate

that developed material.

The final result therefore cannot always be described responsibly by assigning a single origin to the entire unit.

16.16 The Case as a Model of Intellectual Provenance

The episode provides perhaps the clearest compact model of the broader human–AI production process documented in this record:

human intellectual intervention

→

substantial AI analytical expansion

→

AI expressive realization

→

human evaluation and editorial integration.

This model does not apply identically to every passage in the book.

Its importance is evidentiary.

It proves that such a layered provenance sequence occurred and can be reconstructed from surviving records.

16.17 “Right to Be Wrong” Finding

The evidence supports the following attribution:

GOVERNING CONCEPT: HUMAN

ANALYTICAL DEVELOPMENT: AI / ITERATIVE

EXPRESSIVE FORMULATION: AI

EDITORIAL INTEGRATION: HUMAN WITHIN AN ITERATIVE PROCESS.

This finding illustrates a central principle of the dossier:

the identity of the participant who generated the final sentence does not, by itself, determine the provenance of the concept embodied in that sentence; and the identity of the participant who originated the concept does not, by itself, determine the provenance of the resulting analytical development or expression.

For that reason, the “Right to Be Wrong” case serves as one of the strongest concrete demonstrations of layered intellectual provenance in Why Trust Oreskes’s Plumber?.

17. AI-Originating Developments

A complete provenance record must identify not only human-originating concepts subsequently developed by ChatGPT, but also developments for which the surviving record supports a more direct AI-originating contribution.

The term AI-originating development is used here cautiously.

It does not mean that the broader intellectual problem, project, or governing framework originated with ChatGPT.

It refers to a narrower class of contributions in which ChatGPT appears to have introduced a consequential:

• architecture;
• analytical extension;
• structural solution;
• conceptual formulation;
• argumentative development;
• placement proposal;
• analogy;
• or expressive formulation

that was not exhaustively contained in the immediately preceding human input.

Some such contributions were later modified, redirected, or rejected.

Others survived substantially into the developing or canonical manuscript.

This section therefore addresses AI initiative at the level of contribution, not autonomous authorship of the work.

17.1 Initial Seventeen-Chapter Architecture

The clearest early example is the seventeen-chapter book architecture proposed on 27 May 2026.

The intellectual problem and critical orientation of the project had already been established.

ChatGPT nevertheless made a substantial additional contribution by converting that developing field of inquiry into a book-length sequence.

The architecture organized material involving:

• Oreskes;
• scientific trust;
• consensus;
• philosophical assumptions;
• historical examples;
• peer review;
• technological achievement;
• contested inference;
• contemporary problems in science;
• and public trust.

The surviving record attributes this initial macro-architectural organization substantially to ChatGPT.

Capitanu accepted it as the working structure and subsequently modified and directed its development.

The appropriate provenance description is therefore:

intellectual project — HUMAN ORIGIN

initial seventeen-chapter macro-architecture — SUBSTANTIAL AI-ORIGINATING CONTRIBUTION

later architectural development — ITERATIVE UNDER HUMAN DIRECTION.

17.2 Architectural Proposals Below the Book Level

ChatGPT also originated or substantially proposed chapter- and section-level structures.

These included:

• possible section sequences;
• internal chapter architectures;
• transitions;
• relationships among arguments;
• and placement proposals.

Not all survived.

The evidentiary importance lies precisely in the fact that they were proposals rather than automatic components of the book.

Their subsequent acceptance, modification, relocation, or rejection permits the record to distinguish:

AI structural initiative

from

human architectural authority.

17.3 The Proposed Placement of “Inspectable Trust”

The development of “inspectable trust” illustrates this distinction.

Capitanu introduced the underlying philosophical question concerning charity versus inspection.

ChatGPT developed the concept and proposed a section built around it.

ChatGPT also proposed placing that material in Chapter 17.

Capitanu rejected that placement and moved the material to Chapter 15 in order to preserve Chapter 17 as the concluding synthesis.

Accordingly:

initiating question — HUMAN

conceptual development and formulation — AI / ITERATIVE

initial placement proposal — AI

final architectural placement — HUMAN.

The placement proposal itself is therefore an identifiable AI-originating structural contribution even though it was not retained.

17.4 Analytical Extensions in Chapter 11

The Chapter 11 audit identifies several developments that were not exhaustively specified in the human input and were substantially developed or introduced by ChatGPT during generation.

These included:

• examination before dismissal;
• a threshold of minimum seriousness;
• scientific borders as fallible rather than automatically self-validating;
• the distinction between refuting a claim and excluding it by classification;
• error as epistemically productive when exposed to correction;
• curricular distinctions;
• and symmetrical inspection.

Some of these concepts developed from prior human constraints.

Others appear as AI analytical extensions of those constraints.

They should therefore be described as:

AI-originating or substantially AI-developed analytical developments within an established human-directed framework.

This formulation is deliberately narrower than saying that ChatGPT originated the governing problem.

17.5 Minimum Seriousness

The idea of a threshold of minimum seriousness is a particularly useful example.

The pre-existing human problem concerned whether controversial claims can be excluded before examination.

ChatGPT developed the further qualification that not every asserted claim acquires an automatic right to scientific consideration.

Some threshold must separate:

• a claim sufficiently serious to warrant examination;
• from propositions that do not yet meet minimal evidentiary or argumentative standards.

This development prevented the broader argument from collapsing into indiscriminate permissiveness.

Within the recovered Chapter 11 development record, it is best classified as a substantive AI analytical contribution.

17.6 Fallible Scientific Borders

Another AI-developed extension concerned scientific boundaries themselves.

The earlier project had already established that methodological rules should remain inspectable.

ChatGPT developed this into the more explicit idea that scientific borders should not be treated as automatically self-validating.

A boundary may be necessary.

It may also be fallible.

The existence of a disciplinary rule does not, by itself, establish that the rule is epistemically optimal in every case.

This was a substantive analytical development within the established human requirement that methodological naturalism remain distinguishable from ontological naturalism.

17.7 Symmetrical Inspection

ChatGPT also substantially developed the principle of symmetry.

By early July, the argument was being articulated in terms of symmetrical respect under symmetrical inspection.

The principle required scrutiny in both directions.

Naturalistic explanations were not to receive automatic insulation from criticism because of institutional or methodological status.

Design arguments were not to receive evidentiary privilege merely because their prior exclusion had been questioned.

The symmetry concerned:

the standard of inspection

rather than

the equality of competing conclusions.

This development was substantially AI-originating at the level of articulation and analytical expansion, while remaining governed by the previously established human constraint of suspended verdict.

17.8 Productive Error

The developing Chapter 11 argument also included the idea that error can be epistemically productive if it remains exposed to correction.

The governing “right to be wrong” principle originated with Capitanu.

ChatGPT substantially developed its consequences.

One such consequence was the idea that a false claim may still contribute to knowledge when:

• it is examinable;
• its weaknesses can be identified;
• and its failure can refine subsequent inquiry.

This should not be classified as wholly AI-originating at the conceptual root.

It is better classified as:

HUMAN-ORIGINATING PRINCIPLE → AI-ORIGINATING / AI-DEVELOPED CONSEQUENCE.

17.9 “Inspectable Trust”

The phrase and developed concept of inspectable trust also illustrate layered provenance.

The initiating question concerning whether science requires philosophical charity or inspection originated with Capitanu.

ChatGPT then developed the concept into a more explicit framework and formulation.

“Inspectable trust” subsequently became one of the central formulations through which the book described warranted public trust.

The correct attribution is therefore not:

AI-originating project concept.

It is:

human-originating problem → substantial AI conceptual and expressive development.

This distinction is important because §17 is not intended to inflate AI-originating contribution by reclassifying AI development of human ideas as wholly AI-originating.

17.10 From “Meteors” to a Research Program

A similar distinction applies to the research-program framework.

Capitanu introduced the “meteors” analogy and asked what Intelligent Design would have to achieve to become scientifically competitive.

ChatGPT then developed that question into a structured framework involving:

• discriminating predictions;
• falsifiability;
• cumulative research performance;
• comparative performance against alternative explanations;
• and avoidance of mere reliance upon explanatory gaps.

The originating problem was human.

The architecture of the response and several of its criteria were substantially AI-developed.

This therefore represents:

human conceptual initiation → substantial AI-originating analytical structure → later manuscript integration.

17.11 Institutional-Mediation Chain

The same pattern appears in the development of the civic extension of the book.

Capitanu introduced COVID-19 as an example of the pathway through which scientific findings become consequential for citizens through institutions, private interests, and governmental authority.

ChatGPT then systematized the concern into a more explicit sequence distinguishing:

evidence

→

interpretation

→

institutional judgment

→

private interests

→

policy

→

public authority.

The civic problem itself was human-originating.

The explicit chain is best treated as a substantial AI analytical development.

17.12 Counterarguments and Adversarial Extensions

ChatGPT also introduced counterarguments and possible objections that were not always specified beforehand.

These functions included:

• identifying alternative explanations;
• proposing weaknesses in the developing argument;
• challenging evidentiary sufficiency;
• distinguishing procedural from substantive claims;
• and generating adversarial hypotheses.

Where such objections materially altered the analysis, they constitute AI-originating intellectual contribution at the level of critique.

Their subsequent treatment remained subject to human evaluation and project-level constraints.

17.13 Expressive Contributions

The record contains clear AI-originating expressive material.

Examples include:

• specific formulations;
• analogies;
• transitions;
• satirical lines;
• and extended prose.

The “right to be wrong” case demonstrates this directly.

The governing concept originated with Capitanu.

The representative sentence expressing the distinction between the right to be wrong and the right to be believed was generated by ChatGPT.

The plumber-related formulation in Chapter 11 provides another identifiable AI expressive contribution.

Such cases must be recognized directly as:

AI-GENERATED EXPRESSION.

17.14 AI Development That Did Not Survive

AI-originating contribution is not limited to material retained in the final book.

Proposed structures, placements, arguments, or formulations that were later rejected remain part of the production record.

Their evidentiary relevance lies in showing:

• what ChatGPT proposed;
• what alternatives were available;
• and where human redirection occurred.

A rejected AI proposal remains an AI contribution to the developmental process even if it does not appear in the canonical text.

17.15 Survival Is Not Required for Provenance

This distinction prevents a common retrospective error.

If provenance were reconstructed only from the final manuscript, rejected AI-originating developments would disappear.

The result would exaggerate human control while underrepresenting the intellectual range of AI participation during development.

The archive therefore treats both:

retained contributions

and

rejected or superseded contributions

as relevant to understanding the production process.

17.16 AI Initiative Does Not Equal Autonomous Creation

The presence of AI-originating developments does not establish that ChatGPT autonomously created the work.

AI initiative occurred within:

• an existing intellectual project;
• accumulated source materials;
• standing constraints;
• pre-drafting structures;
• human validation requirements;
• and continuing editorial authority.

The proper conclusion is therefore not:

ChatGPT independently authored the book.

It is:

ChatGPT sometimes introduced consequential developments that were not exhaustively specified in prior human input and that materially influenced the architecture, analysis, argumentation, or expression of the work.

17.17 AI-Originating Developments Finding

The surviving evidence supports the existence of genuine AI-originating contribution at multiple levels.

These include, most clearly:

• the initial seventeen-chapter macro-architecture;
• chapter and section structural proposals;
• particular analytical extensions within Chapter 11;
• specific structural placements;
• counterarguments and adversarial developments;
• developed criteria within larger human-originating questions;
• and substantial expressive formulations.

Other cases are more accurately described as AI-developed rather than AI-originating because their governing concept is documented as human.

The distinction is essential.

The provenance record therefore recognizes:

AI ORIGIN

where the record supports AI introduction of a consequential element;

AI DEVELOPMENT

where ChatGPT substantially elaborated a prior human concept or constraint;

and

ITERATIVE DEVELOPMENT

where the resulting contribution cannot accurately be reduced to either participant acting independently.

This framework prevents both artificial inflation and artificial minimization of AI contribution.

18. Five Criteria for AI-Enhanced Authorship

The expression AI-enhanced authorship is used in this record as a descriptive category for the documented production process.

It is not asserted as:

• a statutory category;
• a judicial test;
• an established Canadian legal doctrine;
• a determination that ChatGPT is a legal author;
• or a determination of the copyright status of particular passages.

The term is useful only if it distinguishes a particular production process from both:

• incidental AI assistance;
• and substantially autonomous AI production concealed behind human ownership or approval.

For that reason, the classification is tested through five cumulative criteria.

All five must be satisfied.

Failure to establish any one of them requires reconsideration, qualification, or abandonment of the classification.

18.1 Why the Criteria Are Cumulative

The criteria are designed to prevent several different forms of misclassification.

If substantial AI contribution were absent, the word AI-enhanced would exaggerate the role of artificial intelligence.

If human intellectual origin were absent, the label could conceal an AI-originating work behind a nominal human commissioner.

If continuing human direction and substantive human intellectual contribution were absent, repeated prompting and later selection might be mistaken for meaningful human authorship.

If final human editorial responsibility were absent, the resulting production relationship would require a different analysis.

The classification therefore depends upon the combination of all five conditions.

18.2 Criterion One — Human Intellectual Origin

Finding: SATISFIED

The first criterion asks whether the work possesses a meaningful human intellectual origin.

The relevant question is not whether every later idea originated with the human participant.

It is whether:

• the project;
• problem;
• thesis;
• conceptual framework;
• or other substantial originating elements

can be attributed to human intellectual activity before substantial AI development began.

A general instruction such as:

“Create a book about science and public trust”

would not, by itself, satisfy the criterion.

Application to Why Trust Oreskes’s Plumber?

The surviving record supports a human intellectual origin.

The project did not begin as a general request that ChatGPT devise a book.

Its initiating problem concerned the grounds on which the public is asked to trust science and expertise, particularly the difference between:

being instructed to trust

and

being able to inspect the mechanisms that might warrant that trust.

The plumber analogy became the vehicle through which this problem was developed.

The early project already included questions concerning:

• consensus;
• methodological rules;
• institutional authority;
• peer review;
• historical failures;
• conflicts of interest;
• scientific fallibility;
• controversial inference;
• public policy;
• and warranted trust.

Capitanu also established governing constraints concerning:

• tone;
• fairness to opposing positions;
• methodological versus ontological questions;
• and the treatment of Intelligent Design.

ChatGPT subsequently made major architectural and analytical contributions.

The initial seventeen-chapter structure itself was a substantial AI contribution.

That does not negate the prior human intellectual problem.

The criterion therefore supports:

HUMAN INTELLECTUAL ORIGIN + SUBSTANTIAL SUBSEQUENT AI DEVELOPMENT.

Criterion One is satisfied.

18.3 Criterion Two — Continuing Human Direction

Finding: SATISFIED — STRONGLY DOCUMENTED

The second criterion requires human involvement to continue beyond initiation.

Human direction must be meaningful and consequential.

It may concern:

• objectives;
• boundaries;
• sources;
• arguments;
• architecture;
• research;
• revision;
• or permissible conclusions.

Continuity is evaluated from the developmental record.

It is not inferred from:

• prompt count;
• conversation length;
• frequency of interaction;
• or mere persistence of human participation.

Application to Why Trust Oreskes’s Plumber?

The record contains repeated evidence that human direction continued throughout the project.

Capitanu made consequential decisions concerning:

• which chapters were authorized for drafting;
• which issues were premature;
• how Oreskes and other positions should be represented;
• preservation of conceptual distinctions;
• treatment of Intelligent Design;
• the suspended verdict;
• placement of arguments;
• chapter relationships;
• requirements for further research;
• and later revision.

The pre-drafting protocol is particularly important.

Development was repeatedly mediated through:

• continuity documents;
• Analytical Dossiers;
• source materials;
• manuscript-state information;
• page strategy;
• evidentiary requirements;
• and validation before drafting.

ChatGPT exercised substantial analytical initiative within this system.

It did not possess independent authority to determine the final intellectual direction of the book.

This distinguishes the project from a process in which a human initiates a task and then becomes primarily a recipient of autonomous AI production.

Criterion Two is satisfied and strongly documented.

18.4 Criterion Three — Substantive Human Intellectual Contribution

Finding: SATISFIED

The third criterion is designed to distinguish substantive human intellectual participation from:

• prompting;
• approval;
• selection;
• ownership;
• or mechanical modification of generated output.

Relevant evidence may include:

• introduction or material development of concepts;
• formulation of research questions;
• source selection;
• direction of preparatory documentation;
• definition of chapter functions;
• argumentative constraints;
• comparison of alternatives;
• restructuring;
• substantive revision;
• rejection and redirection;
• and consequential selection and arrangement.

Contemporaneous evidence of intellectual organization before prose generation is particularly relevant.

Application to Why Trust Oreskes’s Plumber?

The record supports substantive human contribution beyond minimal participation.

Representative examples include:

• the initiating problem of warranted and inspectable scientific trust;
• substantive use of the plumber analogy as an epistemological device;
• insistence upon methodological versus ontological naturalism;
• constraints preventing premature Intelligent Design treatment;
• use of Intelligent Design as an admissibility stress test rather than predetermined verdict;
• the distinction between procedural admissibility and epistemic success;
• the suspended-verdict requirement;
• introduction of the “right to be wrong” principle;
• decisions concerning Chapters 11 and 12;
• relocation of important material when AI-proposed placement conflicted with the desired architecture;
• introduction of the institutional-mediation problem;
• substantive corrections preventing oversimplification;
• and consequential decisions concerning the book’s final intellectual and narrative organization.

The Chapter 11 reconstruction provides the strongest granular demonstration.

It shows human intellectual inputs before prose generation, continuing during development, and constraining the permissible conclusion of AI-generated material.

The criterion is not satisfied merely because Capitanu:

• wrote prompts;
• chose among alternatives;
• owned the project;
• or performed final proofreading.

It is satisfied because the record documents:

problem formation, conceptual distinction, argumentative constraint, structural decision-making, substantive correction, and intellectual synthesis.

The evidence does not show that every major idea originated with Capitanu.

That is not required.

Criterion Three is satisfied.

18.5 Criterion Four — Substantive AI Contribution

Finding: SATISFIED — UNEQUIVOCALLY

The fourth criterion requires AI participation itself to be substantial.

Incidental uses such as:

• spelling correction;
• formatting;
• isolated brainstorming;
• or occasional proofreading

would not justify the stronger descriptive category.

Relevant AI participation includes material contribution to:

• research;
• analysis;
• architecture;
• argumentation;
• drafting;
• critical testing;
• rewriting;
• or revision.

Application to Why Trust Oreskes’s Plumber?

There is no serious evidentiary basis for describing ChatGPT’s participation as incidental.

ChatGPT contributed substantially to:

• research assistance;
• source identification and verification;
• organization of research material;
• chapter and section architecture;
• analytical development;
• generation and testing of arguments;
• counterarguments;
• identification of weaknesses and inconsistencies;
• restructuring;
• drafting and redrafting;
• formulations;
• analogies;
• transitions;
• substantial prose;
• continuity management;
• bibliography and footnote-related work;
• adversarial testing;
• editorial revision;
• and working documentation.

The Chapter 11 reconstruction demonstrates that some AI contributions were not merely mechanical execution of fully specified human instructions.

ChatGPT sometimes introduced:

• analytical developments;
• structural solutions;
• formulations;
• and expressive material

that were not exhaustively contained in prior human input.

Some of these survived into the developing manuscript.

Descriptions such as:

“AI-assisted proofreading,”

“AI copyediting,”

or

“AI used only for research”

would therefore materially understate the documented role.

At the same time, substantial AI contribution does not itself establish:

• legal AI authorship;
• legal co-authorship;
• or autonomous AI creation of the work.

Criterion Four is satisfied unequivocally.

18.6 Criterion Five — Human Editorial Responsibility and Control

Finding: SATISFIED — STRONGLY DOCUMENTED AT PROJECT LEVEL

The fifth criterion requires a human participant to determine what is ultimately incorporated into and published as the work and to accept responsibility for the published form.

This includes authority to:

• accept;
• reject;
• redirect;
• revise;
• reorganize;
• or remove AI contributions.

Final approval does not retroactively convert AI-generated expression into human-generated expression.

Editorial control and textual provenance remain distinct.

Application to Why Trust Oreskes’s Plumber?

The record supports continuing and final human editorial authority.

Capitanu determined:

• whether drafting could proceed;
• whether arguments required correction;
• whether structures were retained or replaced;
• whether conclusions exceeded the available evidentiary warrant;
• whether material entered the manuscript;
• and when chapters and production versions became canonical.

Examples include:

• rejection or redirection of proposed structures;
• preservation of conceptual distinctions;
• relocation of material between chapters;
• requirements for additional nuance or evidence;
• maintenance of the suspended verdict;
• and later canonicalization and version control.

The publication stage supplies additional evidence of final control.

However:

control over publication is not equivalent to authorship.

Its significance arises from the fact that publication control forms part of a much larger documented history of intellectual and editorial direction.

ChatGPT materially influenced the pool of available content and generated substantial material retained in the work.

It did not possess independent authority to:

• publish;
• canonize;
• or determine the final state of the manuscript.

Criterion Five is satisfied and strongly documented at project level.

18.7 Cumulative Assessment

The five criteria produce the following result:

Criterion 1 — Human Intellectual Origin SATISFIED.

Criterion 2 — Continuing Human Direction SATISFIED — strongly documented.

Criterion 3 — Substantive Human Intellectual Contribution SATISFIED.

Criterion 4 — Substantive AI Contribution SATISFIED — unequivocally.

Criterion 5 — Human Editorial Responsibility and Control SATISFIED — strongly documented at project level.

The significance of this result is cumulative.

No single criterion would be sufficient.

Human intellectual origin alone would not distinguish the work from a project in which the originating person delegates almost all consequential creation to AI.

Human editorial control alone would not establish intellectual contribution.

Substantive AI contribution alone would not establish a human-directed authorship process.

The documented combination is what matters.

18.8 What the Five-Criterion Finding Means

The criteria support a descriptive conclusion about the developmental process.

They indicate that the work:

• originated in meaningful human intellectual agency;
• remained under continuing human direction;
• contains substantive human intellectual contribution;
• was materially developed and expressed through substantial AI contribution;
• and remained subject to final human editorial responsibility and control.

Within this dossier, that combination supports the descriptive category:

AI-enhanced authorship.

18.9 What the Five-Criterion Finding Does Not Mean

Satisfaction of the criteria does not establish that:

• every passage is legally attributable to the human author;
• ChatGPT is a legal co-author;
• AI-generated text ceases to be AI-generated because it was selected by a human;
• the five criteria have statutory or judicial force;
• or the copyright status of particular passages has been resolved.

The test is evidentiary and descriptive.

It characterizes the documented production process.

18.10 Falsifiability of the Classification

The criteria were developed so that the classification could fail.

If the evidence had shown that:

• the intellectual project itself was substantially devised by AI from a general human request;
• human participation consisted principally of prompting and approval;
• substantive intellectual decisions were delegated to ChatGPT;
• or final editorial responsibility was not meaningfully exercised,

the classification would have required qualification or rejection.

Likewise, if AI participation had been limited to incidental proofreading or formatting, AI-enhanced would have exaggerated its role.

The fact that all five criteria are supported is therefore an evidentiary result, not a premise assumed at the outset.

18.11 Evidentiary Basis

The five-criterion assessment rests upon multiple forms of evidence considered together:

• early project-development records;
• longitudinal human–AI conversations;
• repeated pre-drafting checkpoints;
• research and planning documentation;
• human constraints and redirections;
• substantial AI architecture and drafting;
• manuscript checkpoints;
• the granular Chapter 11 reconstruction;
• version-control evidence;
• and canonical production records.

The result does not depend upon assigning provenance to every sentence in the book.

It depends upon establishing the character of the production process with reasonable evidentiary sufficiency.

18.12 Five-Criteria Finding

On the evidentiary framework adopted in this dossier, Why Trust Oreskes’s Plumber? satisfies all five cumulative criteria proposed for the descriptive category AI-enhanced authorship.

The finding rests upon two propositions supported simultaneously by the surviving record:

the work originated in and remained subject to consequential human intellectual agency;

and

artificial intelligence contributed substantially to its intellectual development and expression.

Neither proposition is sufficient without the other.

Their documented combination is the basis for the classification.

19. Alternative Descriptions

The evidentiary reconstruction permits several simpler descriptions of the production of Why Trust Oreskes’s Plumber? to be tested against the surviving record.

The purpose of this exercise is not to identify the most rhetorically favorable description.

It is to determine which account most accurately preserves the documented contributions of both Daniel Capitanu and OpenAI’s ChatGPT without artificially enlarging or diminishing either.

A satisfactory description must account simultaneously for:

• human intellectual origin;
• continuing human direction;
• substantive human intellectual contribution;
• substantial AI architecture and analysis;
• substantial AI drafting and expression;
• iterative development;
• rejection and redirection;
• final human editorial control;
• and the mixed provenance of some resulting material.

No description that omits one of these major dimensions can adequately represent the documented production process.

19.1 Alternative A: “The Book Was Written by AI from a Human Idea”

This description captures one genuine feature of the record:

ChatGPT generated substantial prose and made substantial analytical, structural, and developmental contributions.

It therefore has greater descriptive value than claims that AI participation was merely editorial.

It is nevertheless inadequate.

The surviving record does not show a process in which Capitanu:

• supplied a general idea;
• delegated the intellectual realization of that idea to ChatGPT;
• and subsequently selected or approved an essentially autonomous AI production.

Such a description omits documented human intellectual activity occurring before, during, and after AI generation.

That activity includes:

• formation of the initiating trust problem;
• definition of the book’s core intellectual framework;
• argumentative constraints;
• the methodological/ontological naturalism distinction;
• the Intelligent Design stress-test function;
• the suspended-verdict requirement;
• the “right to be wrong” principle;
• research direction;
• decisions concerning chapter relationships;
• rejection and relocation of proposed material;
• substantive correction;
• and final canonical control.

It also fails to account for the systematic pre-drafting process.

The human role cannot therefore be reduced to:

idea provider + final selector.

Finding

NOT SUPPORTED AS AN ADEQUATE DESCRIPTION OF THE DOCUMENTED PROCESS.

The description acknowledges substantial AI generation but materially understates continuing human intellectual agency.

19.2 Alternative B: “Daniel Capitanu Wrote the Book and ChatGPT Merely Assisted with Editing”

This description preserves human intellectual agency.

It fails in the opposite direction.

The documentary record establishes that ChatGPT did substantially more than:

• proofreading;
• copyediting;
• stylistic cleanup;
• formatting;
• or post-draft revision.

Its contributions included:

• the initial seventeen-chapter macro-architecture;
• chapter and section architectures;
• research organization;
• source identification and verification;
• preparation of working documentation;
• conceptual development;
• argumentative development;
• counterargument generation;
• adversarial testing;
• structural diagnosis;
• restructuring;
• drafting;
• redrafting;
• generation of substantial prose;
• formulations;
• analogies;
• transitions;
• and other expressive material.

The Chapter 11 reconstruction is particularly incompatible with the “mere editing” description.

There, ChatGPT contributed materially to:

• argumentative architecture;
• analytical extensions;
• conceptual relationships;
• formulations;
• and prose generation.

Some developments were not exhaustively specified in preceding human input.

Calling this merely “editing” would therefore erase a substantial part of the documented record.

Finding

NOT SUPPORTED.

The description materially understates AI participation in architecture, analysis, argumentation, and expression.

19.3 Alternative C: “The Book Was AI-Generated but Human-Directed”

This formulation improves upon Alternative A because it recognizes continuing human direction.

It remains incomplete.

The term AI-generated, when applied to the entire work, can obscure the fact that generation occurred at multiple levels and within substantial pre-existing human intellectual and documentary structures.

Some prose was directly AI-generated.

Some concepts were human-originating.

Some arguments were substantially AI-developed.

Some architectures were AI-proposed.

Some developments were iterative.

Some text underwent later human redirection or revision.

The book as a whole therefore cannot be reduced cleanly to the provenance category of the immediate generator of much of its prose.

Finding

PARTIALLY DESCRIPTIVE BUT INADEQUATE AS A GLOBAL ACCOUNT.

It captures substantial textual generation while obscuring layered intellectual provenance.

19.4 Alternative D: “The Book Was Human-Written with Extensive AI Assistance”

This formulation also captures part of the process.

The project was human-originating and remained under continuing human direction and editorial control.

But the expression human-written may imply that the human participant generated the underlying prose and that AI functioned primarily as support.

That implication is not supported.

ChatGPT generated substantial original expression capable of entering the manuscript.

The record therefore should not use a formulation that erases or euphemizes direct AI drafting.

Finding

INADEQUATE IF “HUMAN-WRITTEN” IS UNDERSTOOD AS TEXTUAL GENERATION BY THE HUMAN AUTHOR.

The work was human-directed, but substantial expression was AI-generated.

19.5 Alternative E: “ChatGPT Was a Literary Assistant”

The literary-assistant analogy is useful but incomplete.

Some ChatGPT functions have close analogues in conventional human assistance:

• research organization;
• source assistance;
• continuity management;
• fact-checking;
• bibliographic support;
• analytical testing;
• editorial diagnosis;
• and preparation of working documentation.

Other functions resemble developmental or intellectual collaboration:

• proposing architectures;
• expanding arguments;
• identifying implications;
• generating counterarguments;
• reorganizing conceptual sequences;
• and converting human-defined objectives into developed plans.

Still other functions exceed what can responsibly be concealed within the general label assistant:

• substantial generation of potentially publishable prose;
• original formulations;
• analogies;
• and other expressive material.

The most accurate functional description of ChatGPT is therefore plural.

At different stages, it operated analogously to combinations of:

• research assistant;
• analytical assistant;
• developmental editor;
• literary assistant;
• critical interlocutor;
• structural planner;
• and substantial drafting instrument or collaborator.

Finding

USEFUL AS A PARTIAL FUNCTIONAL ANALOGY, BUT INSUFFICIENT AS A COMPLETE DESCRIPTION.

It explains some forms of contribution but cannot account adequately for substantial AI-generated expression.

19.6 Alternative F: “Human–AI Co-Authorship”

This description may appear attractive because both participants made substantial contributions.

It is nevertheless avoided in this record.

The term co-authorship carries legal, institutional, and ordinary-language implications that exceed the purpose of the provenance reconstruction.

The dossier does not determine that ChatGPT possesses legal authorship status.

Nor does it establish a symmetrical relationship between the participants.

Decision authority remained asymmetrical.

Capitanu determined:

• the governing purpose of the project;
• permissible conclusions;
• which structures were retained;
• what entered the manuscript;
• and which state became canonical.

ChatGPT made substantial intellectual and expressive contributions but did not possess independent publication or canonical authority.

Finding

NOT ADOPTED.

The term is too legally and institutionally loaded for the descriptive purpose of this dossier.

19.7 Alternative G: “Hybrid Authorship”

The term hybrid authorship can communicate that both human and AI contributions materially shaped the work.

It nevertheless provides little analytical precision.

It does not distinguish:

• origin;
• direction;
• development;
• textual generation;
• decision authority;
• or editorial control.

It could equally describe a work primarily generated by AI with light human intervention or a predominantly human-written work with substantial AI editing.

Finding

DESCRIPTIVELY POSSIBLE BUT ANALYTICALLY UNDER-SPECIFIED.

It does not capture the production structure with sufficient precision.

19.8 Alternative H: “AI-Assisted Authorship”

This description is not false in the broadest possible sense.

Artificial intelligence did assist the project.

The difficulty is that AI-assisted encompasses an extremely wide range of processes, from:

• spelling correction;
• isolated brainstorming;
• and research lookup

to

• sustained analysis;
• architecture;
• argument development;
• substantial drafting;
• and revision.

Applied without further qualification, the expression risks making the documented AI contribution appear more incidental than it was.

Finding

TECHNICALLY POSSIBLE BUT DESCRIPTIVELY TOO WEAK WITHOUT SUBSTANTIAL QUALIFICATION.

The record requires a stronger acknowledgment of AI contribution.

19.9 Best-Supported Description

The description most consistent with the surviving evidence is:

The work developed through sustained human-directed intellectual production in which AI performed substantial assistive, developmental, analytical, and expressive functions.

This formulation preserves the principal findings simultaneously.

It acknowledges that:

• the intellectual project possessed meaningful human origin;
• human direction continued throughout development;
• the human participant made substantive intellectual contributions;
• ChatGPT made substantial contributions of its own;
• ChatGPT generated significant expression;
• some contributions were iterative;
• and final editorial authority remained human.

It does not require fictional symmetry between the participants.

It does not conceal AI-generated expression.

It does not reduce human participation to commissioning, prompting, or approval.

19.10 Relationship to “AI-Enhanced Authorship”

The expression AI-enhanced authorship is the compact descriptive label used in this record for the longer formulation above.

Its meaning is therefore constrained by the documented production model.

It does not mean:

human authorship with trivial AI assistance.

It does not mean:

AI authorship under nominal human supervision.

It refers to a process in which:

meaningful human intellectual agency originates and continues to direct the project

while

artificial intelligence contributes substantially to research, analysis, architecture, argumentation, drafting, critical testing, rewriting, and revision.

19.11 Alternative-Description Finding

The surviving record supports neither pole of the simplest binary.

It does not support:

AI wrote the book; the human supplied the idea.

Nor does it support:

the human wrote the book; AI merely edited it.

The evidence instead supports a layered production history involving:

human intellectual origin and direction

substantial AI intellectual and expressive contribution

iterative development

final human editorial authority.

Any description of the project that removes one of those components materially distorts the documented process.

20. Overall Provenance Finding

The preceding reconstruction permits a general finding concerning the intellectual provenance and production process of Why Trust Oreskes’s Plumber?.

The conclusion is not derived from title-page attribution, retrospective assertion, or final publication control alone.

It rests upon the combined evidentiary record, including:

• early project-development records;
• contemporaneous human–AI conversations;
• repeated pre-drafting checkpoints;
• preparatory research and planning documentation;
• human constraints and conceptual interventions;
• substantial AI architectural and analytical contribution;
• substantial AI drafting and expressive contribution;
• manuscript checkpoints;
• rejection, redirection, and revision;
• canonical version history;
• and the representative granular reconstruction of Chapter 11.

Taken together, these materials support a stable account of the production process.

20.1 Human Intellectual Origin

The project possesses a meaningful human intellectual origin.

Daniel Capitanu established the initiating problem concerning warranted scientific trust and the inspectability of the mechanisms invoked to justify that trust.

He also established the central critical orientation, rhetorical posture, and a substantial part of the core intellectual framework within which the book developed.

This finding does not imply that every argument in the final work originated with him.

The evidence does not support such a claim.

The appropriate formulation is:

human intellectual origin of the project and core intellectual framework.

20.2 Continuing Human Agency

Human intellectual agency did not disappear after initiation.

The record documents continuing decisions concerning:

• research;
• argumentative boundaries;
• conceptual distinctions;
• chapter functions;
• structural relationships;
• permissible conclusions;
• rejected alternatives;
• evidentiary sufficiency;
• and later canonicalization.

Capitanu therefore remained a consequential intellectual participant throughout development.

His role cannot be reduced to:

• topic selection;
• prompting;
• output selection;
• or final approval.

20.3 Substantial AI Contribution

The record simultaneously establishes substantial AI contribution.

ChatGPT participated materially in:

• research assistance;
• source organization;
• book architecture;
• chapter and section architecture;
• analytical development;
• argument generation;
• counterargument generation;
• critical testing;
• restructuring;
• drafting;
• redrafting;
• formulation;
• analogy;
• continuity management;
• and revision.

Some contributions were assistive.

Some were developmental.

Some were clearly generative and expressive.

Some analytical or structural developments originated with ChatGPT rather than being fully specified in prior human instructions.

This contribution must be acknowledged without minimization.

20.4 Layered Provenance

The provenance of the work is therefore layered.

Different aspects of the same unit may have different origins.

For example:

• a concept may be human-originating;
• its analytical expansion may be AI-developed;
• its formulation may be AI-generated;
• its placement may be human-determined;
• and its final canonical state may result from iterative revision.

The “Right to Be Wrong” case demonstrates this directly.

The broader Chapter 11 reconstruction shows the same pattern at chapter scale.

The final work therefore cannot meaningfully be described through a single variable such as:

• percentage of words;
• percentage of ideas;
• or percentage of labor.

20.5 Iterative Character of the Process

The collaboration was recursive.

It involved repeated sequences of:

human intellectual input

→

AI proposal or development

→

human evaluation

→

acceptance, correction, rejection, or redirection

→

further AI development

→

human integration and control.

This sequence varied across chapters and stages.

The project was neither:

a completed human manuscript subsequently polished by AI

nor

an autonomous AI manuscript produced from a general human request.

20.6 Decision Authority

Decision authority remained human.

ChatGPT could:

• propose;
• analyze;
• draft;
• critique;
• restructure;
• and generate.

It did not independently determine:

• the governing purpose of the work;
• which conclusions were permissible;
• what entered the final manuscript;
• which architecture became canonical;
• or when a publication version was authorized.

This asymmetry is part of the production history.

It does not erase substantial AI contribution.

20.7 Relationship Between Direction and Expression

The record distinguishes human direction from textual generation.

Human editorial and intellectual control does not automatically establish human generation of AI-produced expression.

Likewise, substantial AI generation does not automatically establish AI origin of the concepts, constraints, and research structures embodied in that expression.

These dimensions remain analytically separate.

20.8 Five-Criteria Result

The five cumulative criteria established earlier in this dossier are all supported.

Human Intellectual Origin SATISFIED.

Continuing Human Direction SATISFIED — strongly documented.

Substantive Human Intellectual Contribution SATISFIED.

Substantive AI Contribution SATISFIED — unequivocally.

Human Editorial Responsibility and Control SATISFIED — strongly documented at project level.

This cumulative result provides the immediate evidentiary basis for the overall classification.

20.9 Descriptive Classification

On the evidentiary framework adopted in this dossier, Why Trust Oreskes’s Plumber? satisfies the proposed descriptive classification:

AI-enhanced authorship.

The term is appropriate because the surviving record supports two propositions simultaneously.

First:

the work originated in and remained subject to consequential human intellectual agency.

Second:

artificial intelligence contributed substantially to the development and expression of the work.

Neither proposition adequately describes the project without the other.

20.10 Meaning of the Finding

Within this record, AI-enhanced authorship means that the book developed through:

• meaningful human intellectual origin;
• continuing human direction;
• substantive human intellectual contribution;
• substantial AI assistive, analytical, developmental, and expressive contribution;
• and final human editorial responsibility and control.

The term describes the documented production process.

20.11 What the Finding Does Not Establish

The finding does not mean that:

• every passage is necessarily legally attributable to Daniel Capitanu;
• every AI-generated passage becomes human-generated through selection or approval;
• ChatGPT is designated a legal author or co-author;
• the term AI-enhanced authorship is an established category of copyright law;
• the classification supersedes any AI-disclosure definition required by a publisher, distributor, registration system, platform, legislature, court, or administrative authority;
• or the copyright status of particular AI-generated expression has been resolved.

Those questions remain analytically distinct.

20.12 AI-Generated Material Within an AI-Enhanced Work

The classification also does not replace the category AI-generated where that category applies to particular material.

An AI-enhanced work may contain AI-generated passages.

The descriptions operate at different levels.

AI-generated may describe the immediate origin of a particular expression.

AI-enhanced authorship describes the documented developmental relationship at the level of the work as a whole.

The categories are therefore not mutually exclusive.

20.13 Evidentiary Proportionality

The overall finding does not depend upon reconstructing the provenance of every sentence in the book.

The evidentiary basis instead combines:

• longitudinal evidence across the project;
• contemporaneous documentation;
• repeated human interventions;
• explicit acknowledgment of substantial AI contribution;
• canonical manuscript history;
• and a representative granular chapter audit.

This provides a reasonable basis for characterizing the production process.

The dossier therefore distinguishes between:

sufficient evidence to characterize the production process

and

exhaustive forensic attribution of every component of the work.

The former has been undertaken.

The latter is not claimed.

20.14 Remaining Granular Uncertainty

Some passage-level provenance remains unresolved.

This includes, in particular cases:

• exact sentence-level rewriting;
• complete inventories of rejected generated text;
• first origin of subordinate analytical elements;
• and precise transformation histories between some intermediate drafts and canonical versions.

These unresolved matters do not disappear because the global process can be characterized.

Nor do they prevent a global characterization where longitudinal evidence is sufficiently strong.

20.15 Independence of the Evidentiary Record

The overall finding remains subject to the underlying primary archive.

If future examination of primary evidence contradicts any part of this reconstruction, the appropriate response is to correct the public provenance record.

The archive is not maintained to prove the conclusion.

It is maintained to preserve the developmental history against which the conclusion can be tested.

20.16 Overall Provenance Finding

The most accurate overall reconstruction supported by the surviving record is:

Why Trust Oreskes’s Plumber? originated in a human intellectual project conceived and continually directed by Daniel Capitanu. OpenAI’s ChatGPT made substantial contributions to the research, architecture, analytical and argumentative development, drafting, redrafting, critical testing, revision, and expression through which that project became the resulting work. The collaboration was iterative, with identifiable human-originating, AI-originating, and collaboratively developed contributions. Capitanu retained consequential intellectual direction and final editorial control, while ChatGPT generated substantial analytical and expressive material incorporated into the manuscript.

Within the evidentiary and descriptive framework defined by this dossier, that production history supports the classification:

AI-enhanced authorship.

The classification describes the documented provenance and production process of the work.

It does not purport to resolve the legal authorship or copyright status of every individual component.

21. Limits of the Reconstruction

The reconstruction presented in this record is intended to be rigorous, transparent, and proportionate.

It is not presented as an exhaustive forensic attribution of every intellectual or expressive component of Why Trust Oreskes’s Plumber?.

The distinction is important.

The evidentiary question addressed by this dossier is whether the surviving record is sufficient to characterize the production process responsibly and to support the provenance findings stated above.

That is different from establishing the complete developmental history of:

• every sentence;
• every formulation;
• every source choice;
• every revision;
• every rejected passage;
• every intermediate version;
• and every human or AI intervention across an 823-page work.

The former is the purpose of this record.

The latter is neither claimed nor presently necessary.

21.1 Evidentiary Sufficiency Rather Than Exhaustive Attribution

The overall findings rely upon multiple forms of evidence considered longitudinally.

These include:

• contemporaneous developmental conversations;
• preparatory documentation;
• repeated pre-drafting checkpoints;
• research and planning records;
• manuscript history;
• human interventions and consequential decisions;
• substantial documented AI contribution;
• canonical production history;
• and the representative granular reconstruction of Chapter 11.

Taken together, these materials provide a reasonable evidentiary basis for characterizing the production process.

The dossier therefore distinguishes between:

sufficient evidence to characterize the production process

and

exhaustive forensic attribution of every component of the work.

The existence of the first does not imply completion of the second.

21.2 Uneven Survival of Primary Evidence

The developmental archive is not perfectly uniform.

Some events are preserved through identifiable contemporaneous records.

Others are corroborated through multiple later or parallel records.

Some developmental relationships can only be inferred.

Still others remain unresolved.

The absence of a surviving document does not prove that an event did not occur.

Conversely, later recollection does not become contemporaneous evidence merely because it is plausible or consistent with the final manuscript.

This limitation is addressed through the evidentiary classifications defined earlier:

• DOCUMENTED;
• CORROBORATED;
• INFERRED;
• and UNRESOLVED.

21.3 Incomplete First-Origin Chronology

For some concepts, structures, or formulations, the exact first moment of origin cannot presently be established.

The record may support that a concept existed by a particular date without establishing the precise interaction in which it first appeared.

This applies especially where:

• earlier conversations have not been preserved in complete form;
• intermediate planning documents survive only through references to them;
• or several related developments occurred during an intensive period of iterative work.

Accordingly, this dossier distinguishes between:

first documented appearance

and

demonstrated first intellectual origin.

The former should not automatically be treated as the latter.

21.4 Limits of Retrospective Memory

Retrospective recollection may assist in identifying:

• relevant events;
• missing documents;
• possible dates;
• or questions requiring further investigation.

It receives less evidentiary weight than contemporaneous documentation.

Where a retrospective account cannot be independently supported, the record does not silently elevate it into primary evidence.

This limitation applies even where the retrospective account would favor the author’s provenance claim.

21.5 Passage-Level Textual Provenance

The record does not reconstruct the exact textual provenance of every passage.

For some material, the developmental record permits relatively clear findings concerning:

• AI generation;
• human conceptual origin;
• human revision;
• or iterative development.

For other passages, the exact sequence between:

• initial generation;
• redrafting;
• direct human rewriting;
• AI rewriting;
• selection;
• and final canonical form

has not been reconstructed at sentence level.

Where such information remains unavailable, the dossier does not infer it retrospectively.

21.6 Direct Human Rewriting

One recurring unresolved question concerns the extent of direct human rewriting of AI-generated prose.

The evidence establishes substantive human intellectual and editorial intervention.

It does not always establish:

• which sentences were directly rewritten by Capitanu;
• how extensively particular AI-generated passages were altered;
• or whether the final wording of every retained passage remained primarily AI-generated.

This limitation is particularly explicit in the Chapter 11 audit.

The provenance of concepts, constraints, and decisions may therefore be better established than the exact textual history of every sentence.

21.7 Rejected and Superseded AI Material

The surviving record also does not contain a complete inventory of every AI-generated proposal that was rejected, replaced, or superseded.

Known cases of:

• rejected architectures;
• relocated material;
• abandoned formulations;
• and redirected arguments

demonstrate that rejection occurred.

The full quantitative extent of rejected AI-generated material is not established.

Accordingly, the dossier does not attempt to infer a rejection rate or percentage of AI material ultimately retained.

21.8 Exact Quantification of Contribution

No attempt is made to assign percentages of contribution.

The evidence does not justify a meaningful division such as:

• percentage human;
• percentage AI;
• percentage of words;
• percentage of ideas;
• percentage of intellectual labor;
• or percentage of authorship.

Such measurements would collapse qualitatively different functions into a single numerical scale.

A concept, architecture, formulation, revision, and final editorial decision are not commensurable units.

The iterative production history is therefore described functionally and developmentally rather than numerically.

21.9 Limits of the Representative Chapter Audit

Chapter 11 was selected because it provides unusually rich evidence for examining:

• pre-drafting intellectual architecture;
• human constraints;
• AI analytical development;
• prose generation;
• continuing human intervention;
• and canonical integration.

It is representative of important features of the project.

It is not presumed to prove that every other chapter followed exactly the same sequence.

The broader longitudinal evidence is therefore necessary to determine whether the pattern is project-wide.

The Chapter 11 audit demonstrates a production mechanism.

It does not substitute for every possible chapter-level reconstruction.

21.10 Limits of Functional Analogies

Descriptions such as:

• research assistant;
• literary assistant;
• developmental editor;
• drafting collaborator;
• or analytical assistant

are functional analogies.

They are useful for understanding the types of work performed by ChatGPT.

They do not establish that ChatGPT possessed:

• human legal personality;
• employment status;
• contractual status;
• or legal authorship.

Likewise, no single analogy captures the system’s complete role.

21.11 Limits of the Five-Criteria Framework

The five criteria for AI-enhanced authorship are an evidentiary and descriptive framework developed for this dossier.

They are not presented as:

• statutory requirements;
• judicial doctrine;
• an official copyright test;
• or a universally applicable taxonomy of human–AI authorship.

Their purpose is narrower.

They provide a falsifiable method for determining whether the descriptive term used in this project accurately fits its documented production history.

21.12 Legal Questions Remain Separate

This record does not determine:

• legal authorship of every passage;
• legal co-authorship;
• copyright ownership;
• copyrightability of particular AI-generated expression;
• or the legal consequences of human selection, revision, or editorial control.

Those questions may depend upon:

• applicable law;
• institutional rules;
• contractual arrangements;
• jurisdiction;
• and evidence examined at greater granularity.

The dossier reconstructs intellectual provenance and production history.

It does not substitute for a legal determination.

21.13 Final Manuscript Attribution Is Not Origin Evidence

The fact that production records identify Daniel Capitanu as author and Les Éditions du Rasoir as publisher establishes how the completed work was represented and controlled for publication.

It does not independently prove the provenance of every intellectual or expressive component.

For that reason, publication metadata is treated as evidence of publication status and control, not as a substitute for developmental evidence.

21.14 Canonical Text Is an Endpoint, Not a Complete History

The final manuscript establishes what survived into the published work.

It does not reveal automatically:

• what alternatives were considered;
• which drafts were rejected;
• how concepts evolved;
• who introduced every intermediate development;
• or how the final wording emerged.

The canonical text is therefore an evidentiary endpoint.

The developmental archive is required to reconstruct the process leading to it.

21.15 Unresolved Matters Are Not Presumptions

Where the evidence is insufficient, the designation UNRESOLVED is not treated as evidence against either human or AI contribution.

It means only that reliable attribution cannot presently be made.

The dossier therefore avoids converting uncertainty into:

• presumptive human origin;
• presumptive AI origin;
• or presumptive collaboration.

The absence of attribution is itself an evidentiary result.

21.16 Possibility of Future Correction

This reconstruction remains corrigible.

If additional primary evidence becomes available and materially changes a provenance finding, the appropriate response is to revise the reconstruction.

The purpose of the record is not to preserve a preferred narrative.

It is to preserve the most accurate account supported by the evidence available.

21.17 Reasonable Stopping Point

The representative Chapter 11 audit established:

1. substantial intellectual and documentary preparation before mature generation;
2. identifiable human intellectual contributions before and during development;
3. substantial and sometimes original AI analytical and expressive contribution;
4. continuing human direction and editorial intervention;
5. and survival of material produced through that interaction into the manuscript.

At that point, additional sentence-by-sentence reconstruction would have produced diminishing evidentiary returns relative to the purpose of Level 3.

The dossier therefore adopts a stopping principle of:

reasonable evidentiary sufficiency rather than exhaustive forensic reconstruction.

21.18 Availability of Deeper Verification

The existence of these limits does not make the public record unverifiable.

More detailed primary materials are preserved separately.

Where a legitimate need for greater granularity arises, the Level 4 evidentiary archive provides a pathway for further examination.

The function of Level 3 is therefore not to anticipate every possible future inquiry.

It is to:

• state the conclusions;
• explain the methodology;
• provide representative evidence;
• identify limitations;
• and preserve a path to deeper verification.

21.19 Limits Finding

The reconstruction is sufficiently supported to characterize the documented human–AI production process of Why Trust Oreskes’s Plumber?.

It is not exhaustive.

Some granular questions concerning:

• first origin;
• sentence-level rewriting;
• rejected material;
• source-selection provenance;
• and intermediate version history

remain partly unresolved.

Those limitations are preserved explicitly rather than silently filled.

They do not negate the longitudinal findings established by the evidence.

They define the proper evidentiary scope within which those findings should be understood.

22. Underlying Evidentiary Archive

The public transparency framework for Why Trust Oreskes’s Plumber? distinguishes between disclosure, explanation, accountability, and verification.

The underlying primary materials belong to the fourth of those functions.

They constitute a restricted evidentiary archive retained to permit further verification of the developmental history described in this record.

The archive is not itself part of the general public record.

Its existence is intended to make the public provenance account verifiable without requiring indiscriminate publication of private developmental materials.

22.1 Principle

Transparency concerning the use of artificial intelligence does not require unrestricted publication of the complete developmental history of the work.

Three obligations must remain distinct:

public disclosure of the collaboration;

public explanation of its intellectual and developmental character;

and

access to underlying primary evidence.

These are not equivalent.

The objective is:

meaningful and verifiable transparency

rather than

unrestricted publication of private working records.

22.2 The Four-Level Transparency Architecture

The project uses four levels.

Level 1 — Notice

The published book contains a concise disclosure identifying material AI involvement and directing interested readers toward fuller information.

Its function is notice.

Level 2 — Explanation

The publisher’s website provides an accessible explanation of:

• the human–AI collaboration;
• the substantial AI role;
• continuing human intellectual direction;
• and the descriptive meaning of AI-enhanced authorship.

Its function is public transparency.

Level 3 — Accountability

The present Human–AI Collaboration and Intellectual Provenance Record provides:

• methodology;
• evidentiary vocabulary;
• developmental chronology;
• competing explanations;
• representative evidence;
• the Chapter 11 reconstruction;
• the five-criterion analysis;
• limitations;
• and the basis for the resulting descriptive classification.

Its function is reasoned public accountability and inspectability.

Level 4 — Verification

The restricted archive preserves underlying primary materials from which more granular examination can be undertaken where legitimately required.

Its function is verification.

The resulting architecture is:

BOOK

→

PUBLIC COLLABORATION STATEMENT

→

PUBLIC PROVENANCE RECORD

→

RESTRICTED EVIDENTIARY ARCHIVE.

22.3 Materials Potentially Included

Depending upon relevance and preservation, Level 4 may include:

• contemporaneous ChatGPT conversations;
• preparatory research documents;
• Continuity Briefs;
• Analytical Dossiers;
• chapter and section planning records;
• source materials used during development;
• manuscript drafts;
• intermediate versions;
• revision histories;
• source-verification records;
• editorial instructions;
• corrections;
• placement and integration records;
• canonical production files;
• version-control records;
• and other materials capable of corroborating or testing the developmental history described in this record.

The list is functional rather than exhaustive.

Not every category necessarily contains a complete record for every stage of the project.

22.4 Purpose of Preservation

The archive is preserved for verification.

Its purpose is to make it possible, where necessary, to test:

• chronology;
• provenance claims;
• developmental sequences;
• version relationships;
• particular attributions;
• or other assertions contained in Level 3.

Preservation does not require publication.

The objective is to preserve the possibility of later examination.

22.5 No General Right of Public Access

The existence of Level 4 does not create a general right of public access to private developmental records.

Public curiosity alone does not establish entitlement.

Requests may be declined where they are:

• insufficiently identified;
• excessively broad;
• unrelated to a legitimate verification question;
• disproportionate to the issue being examined;
• or otherwise inappropriate.

This limitation is not intended to prevent legitimate scrutiny.

It distinguishes:

verifiability

from

unrestricted disclosure.

22.6 Legitimate Verification Requests

Access to relevant Level 4 material may be considered where a sufficiently identified verification need exists.

Examples may include:

• judicial proceedings;
• administrative proceedings;
• copyright or other intellectual-property examination;
• legitimate governmental or regulatory review;
• contractual verification;
• publishing verification;
• serious scholarly research;
• professional or expert examination of intellectual provenance;
• or another reasonably demonstrated legal, academic, institutional, or evidentiary purpose.

The inclusion of a category does not create automatic entitlement.

Relevant considerations may include:

• identity of the requester;
• purpose of the request;
• materials sought;
• specificity of the inquiry;
• and relevance of the requested material.

22.7 Proportionality

Where access is appropriate, disclosure should ordinarily be limited to what is reasonably necessary to examine the particular provenance question or claim at issue.

A request concerning:

• one concept;
• one chapter;
• one attribution;
• or one developmental event

does not ordinarily justify disclosure of the complete archive.

The governing principle is:

disclose what is reasonably necessary to permit meaningful verification of the matter being examined.

22.8 Data Minimization

Proportional verification may therefore take different forms.

Access may involve:

• selected records;
• relevant excerpts;
• specified document sets;
• redacted copies;
• controlled or supervised inspection;
• verification of particular records without unrestricted copying;
• or broader access where the circumstances genuinely require it.

The least extensive form of disclosure capable of permitting meaningful verification should ordinarily be preferred.

22.9 Grounds for Restriction or Redaction

Access may be restricted, conditioned, or redacted where reasonably necessary to protect:

• personal information;
• private communications unrelated to the provenance inquiry;
• confidential information;
• third-party information or rights;
• unpublished intellectual property;
• commercially sensitive material;
• credentials;
• account information;
• security-related information;
• material outside the relevant scope of inquiry;
• or information subject to legitimate legal, contractual, ethical, privacy, or security constraints.

Where feasible, relevant evidentiary material should be separated from unrelated protected material rather than treating the existence of protected information as an automatic reason to deny all verification.

22.10 Conditions of Access

Where appropriate, access may be subject to reasonable safeguards.

These may include:

• identification of the requester;
• identification of the verification purpose;
• specification of the claims or materials at issue;
• restrictions on copying;
• restrictions on redistribution;
• confidentiality undertakings;
• controlled inspection;
• redaction;
• or other proportionate conditions.

Different contexts may justify different access arrangements.

A:

• scholarly inquiry;
• contractual review;
• administrative examination;
• regulatory inquiry;
• and judicial proceeding

need not be handled identically.

22.11 Mandatory Legal or Regulatory Process

The Level 4 access policy governs voluntary disclosure.

It does not purport to override:

• binding legal obligations;
• valid court orders;
• regulatory requirements;
• compulsory process;
• contractual duties;
• or other legally enforceable disclosure requirements.

Where such obligations apply, disclosure is governed by the applicable legal or procedural framework.

22.12 Preservation Standards

Because Level 3 depends upon the existence of an underlying documentary record, reasonable preservation of Level 4 materials is part of the transparency architecture.

Preservation should, where practicable, maintain:

• document identity;
• chronology;
• version relationships;
• original or canonical files;
• relevant metadata where available;
• the distinction between contemporaneous records and later reconstructions;
• and sufficient organization to permit relevant evidence to be located if verification becomes necessary.

Preservation should also avoid unnecessary alteration of primary material.

The archive exists to preserve the historical record, not to improve it retrospectively.

22.13 Evidentiary Independence of Level 4

The restricted archive must remain evidentially independent of Level 3.

Level 4 should not be modified to make primary history conform to the conclusions of the public provenance record.

If a preserved primary record conflicts with a later reconstruction:

the primary record remains preserved in its authentic form

and

the reconstruction should be corrected where necessary.

This principle is fundamental.

The archive does not exist merely to support the public account.

It also exists to test it.

22.14 Archive as a Check on Level 3

A credible evidentiary system must permit its conclusions to be challenged by its own source material.

Accordingly, Level 4 performs two simultaneous functions.

It can:

corroborate Level 3

but it can also:

contradict Level 3.

The second function is as important as the first.

Without it, the archive would function as advocacy material rather than evidence.

22.15 Distinction Between Preservation and Publication

Preservation does not imply:

• public posting;
• open download;
• unrestricted inspection;
• or indefinite access by any requester.

The archive may remain private while still performing a meaningful verification function.

The critical requirement is that relevant primary materials remain sufficiently identifiable and organized to permit examination where a legitimate need arises.

22.16 Scope of Verification

Verification may occur at different levels.

A request may concern:

• whether a particular concept existed before drafting;
• whether ChatGPT generated a specific formulation;
• whether a human redirection occurred;
• whether an intermediate manuscript contained particular material;
• whether a version was canonical;
• or whether the public reconstruction accurately describes a particular sequence.

The scope of access should correspond to the scope of the verification question.

22.17 No Promise of Complete Archive Disclosure

Authorization of access to some Level 4 material does not imply entitlement to the entire archive.

A legitimate inquiry may be answered through:

• a limited document set;
• selected excerpts;
• metadata;
• or controlled inspection.

Broader disclosure may be justified where circumstances require it.

It is not presumed.

22.18 Relationship Between Level 3 and Level 4

Level 3 answers:

What does the surviving evidence support concerning the development of the work?

Level 4 answers:

What primary materials exist from which that reconstruction can be tested more deeply?

The two functions are complementary.

Level 3 provides public interpretive accountability.

Level 4 preserves the evidentiary basis for further verification.

22.19 Relationship Across All Four Levels

The four levels therefore operate at increasing resolution.

LEVEL 1 — NOTICE Material AI involvement is disclosed.

LEVEL 2 — EXPLANATION The nature of the collaboration is explained publicly.

LEVEL 3 — ACCOUNTABILITY The evidentiary reasoning and representative documentation are presented.

LEVEL 4 — VERIFICATION Underlying primary materials are preserved for controlled examination.

The system is designed to avoid two opposite failures.

It does not ask the public to accept an unsupported assertion concerning human–AI provenance.

It also does not treat transparency as requiring publication of every private developmental record associated with the work.

22.20 Public Access Statement

Where a concise public description of Level 4 is required, the following formulation accurately reflects the policy:

Underlying archival materials are retained for verification purposes but are not part of the public record. Access to relevant materials may be granted, where appropriate, in response to a sufficiently identified scholarly, legal, regulatory, contractual, or other legitimate verification need, subject to authorization and any necessary confidentiality, privacy, intellectual-property, security, or third-party restrictions. The existence of the archive does not create a general right of public access to private developmental records.

This statement describes an access policy.

It is not a promise that:

• every request will be granted;
• every relevant record can necessarily be produced;
• or the complete archive will be disclosed whenever access is authorized.

22.21 Archive Finding

The Level 4 archive is an integral part of the transparency architecture because it preserves the distinction between:

public accountability

and

primary evidentiary verification.

Its purpose is neither secrecy nor indiscriminate disclosure.

It is to preserve a sufficiently authentic and organized record from which material claims concerning the development of Why Trust Oreskes’s Plumber? can be tested where legitimate circumstances require deeper examination.

Its governing principle is therefore:

preserve the evidence as evidence.

If the primary archive supports Level 3, it strengthens the public reconstruction.

If it contradicts Level 3, the reconstruction must yield to the evidence.

23. Concluding Statement

This record is intended to provide a transparent and proportionate account of the development of Why Trust Oreskes’s Plumber? and of the substantial roles played by both Daniel Capitanu and OpenAI’s ChatGPT.

The surviving evidence supports a production history in which the conception, intellectual direction, and core intellectual framework of the work originated with Capitanu and were developed under his continuing direction, while ChatGPT contributed substantially to research, structural and argumentative development, analytical testing, drafting, redrafting, formulation, revision, and other parts of the work’s intellectual and expressive development.

The collaboration was iterative. It cannot accurately be reduced either to a completed human manuscript subsequently polished by artificial intelligence or to an autonomously generated AI manuscript produced from a general human request. Human-originating, AI-originating, and collaboratively developed contributions entered the work at different stages and in different forms, with intellectual provenance varying across conception, research, architecture, argument, formulation, drafting, revision, selection, and final editorial control.

The conclusions of this record concern the documented intellectual provenance and production process of the work.

Within that framework, the expression AI-enhanced authorship is used descriptively. It is not asserted as an independent legal category, as a designation of ChatGPT as a legal author or co-author, or as a determination of the copyright status of particular passages or other expressive elements.

The record deliberately preserves the distinction between:

intellectual provenance,

descriptive characterization of the authorship process,

and

legal copyright determination.

Those questions overlap in some respects, but they are not identical.

The reconstruction is also deliberately proportionate rather than exhaustive. It provides sufficient longitudinal and representative evidence to characterize the documented production process without claiming sentence-by-sentence forensic attribution across the entire work.

Where the surviving evidence does not permit reliable attribution, uncertainty remains explicit rather than being resolved retrospectively in favor of either the human or AI participant.

More detailed primary documentation is preserved separately in the restricted evidentiary archive. That archive exists to permit further verification where a legitimate scholarly, legal, regulatory, contractual, professional, or other evidentiary need arises, subject to appropriate access controls and protections.

The archive is independent of the conclusions of this record. If preserved primary evidence materially conflicts with any later reconstruction, the primary evidence should remain intact and the reconstruction should be corrected accordingly.

With these qualifications, the four levels of the transparency architecture should be understood as progressively more detailed descriptions of the same documented human–AI collaboration:

Level 1 — Notice

Level 2 — Explanation

Level 3 — Accountability

Level 4 — Verification

Taken together, they are intended to make the collaboration neither invisible nor overstated, and to permit the provenance claims surrounding Why Trust Oreskes’s Plumber? to remain open to informed inspection.