Abstract
“AI was used in preparing this work.” The sentence appears transparent. It may also tell a reader almost nothing. Did a tool correct grammar, suggest topics, retrieve possible sources, produce a first-pass paragraph, translate an interview, make an image, or help organize research notes? What did the human writer verify, reject, revise, or take responsibility for? What record exists if an editor needs clarification? And which parts of a private creative process should not be exposed to satisfy a generic demand for transparency?
This paper introduces the responsibility boundary: the location where a human actor remains answerable for published expression, claims, source use, and relevant representations even when a system assists with ideation, revision, research, or generation. It separates assistance from authorship, public disclosure from private documentation, and public publishing practice from U.S. copyright-registration requirements. The aim is neither concealment nor a complete surveillance record of creative work.
The paper develops the Responsibility Boundary Canvas, a six-field instrument that asks about assistance type, material effect, human review, audience, retained record, and privacy/confidentiality. Scholarly ethics guidance correctly emphasizes that AI systems cannot assume authorship accountability and that human authors remain responsible for submitted work.[1] But policy examples already differ in their disclosure thresholds, and journalism guidance frames transparency in relation to audience trust, public benefit, human review, accuracy, and privacy.[3] [4] A recent open-science framework offers a further distinction: disclosure communicates that AI was used, while documentation makes workflow traceable; the appropriate record should be task-specific, proportionate, role-specific, and privacy-sensitive.[6] The result is a contextual teaching method, not legal advice, an authorship test, or a universal disclosure statement.
1. The sentence that says almost nothing
A publication carries a note at the end: “AI was used in preparing this work.” A reader sees it after the argument has already made its claims. An editor sees it while deciding whether the note is sufficient. A writer sees it as a demand to be honest. Yet the sentence cannot tell any of them what they actually need to know.
Was the assistance limited to grammar and spelling? Did it generate a working summary of a source that the writer then checked against the original? Did it produce language later revised? Did it influence a substantive interpretation? Did it create an illustration? Did it handle confidential notes, unpublished writing, or another person’s material? Were citations verified? Did a person review the final expression and accept responsibility for it?
The sentence may be accurate while remaining operationally empty. It can also create the opposite problem: a reader may infer that every use of a tool carries the same weight, or that use of a system displaces the person who chose, checked, revised, and published the work.
Disclosure is useful when it helps an accountable human being communicate a material fact to a particular audience. It is not useful when it turns every act of assistance into an undifferentiated confession.
This paper begins from a simple distinction. Assistance is not authorship. A tool may alter a workflow. It does not thereby become able to accept responsibility for accuracy, attribution, confidentiality, rights, publication agreement, or the final form of a work. The Committee on Publication Ethics (COPE) states this directly for research publications: AI tools cannot meet authorship requirements because they cannot take responsibility for submitted work, manage conflicts of interest, or manage copyright and license agreements; human authors remain responsible for content, including AI-produced parts.[1]
That statement offers an important responsibility principle. It does not, by itself, answer every question about essays, journalism, fiction, books, professional reports, or private draft work. The problem is therefore not whether to copy a scholarly policy into every creative context. It is how to separate the categories that a short disclosure sentence often collapses.
2. Six categories that should not collapse
The word disclosure can refer to several different acts. Confusing them produces either misleading minimalism or unnecessary surveillance.
| Category | Working question | Example | What it is not |
|---|---|---|---|
| Assistance | What did the system do? | It suggested alternate phrasing, surfaced possible sources, generated a paragraph, or translated a draft. | A declaration that the system authored, approved, or verified the work. |
| Authorship | Who accepts the role of responsible creator/publisher in the relevant context? | A human author signs, revises, approves, and answers for the published work. | A count of which entity produced the most words. |
| Responsibility | Who checked claims, sources, reasoning, rights, and final form? | A human writer and/or editor verifies, corrects, and accepts consequences. | A responsibility transferred to a tool through a note. |
| Public disclosure | What does a reader or publisher need to know about material assistance? | A concise note explaining a relevant generative or research role. | A full prompt archive. |
| Private documentation | What record supports internal review or later clarification? | A retained source-check note or version record. | Mandatory public exposure of drafts, private notes, or confidential materials. |
| Registration disclosure | What must an applicant disclose to a named authority under its rules? | A U.S. copyright-registration application describes human and AI-generated material. | A general policy for readers, publishers, or every jurisdiction. |
The categories overlap, but they do not replace one another. A publisher can require an author to disclose a material generative use to an editor while allowing a shorter public note. A writer can retain a record of source verification without publishing private prompts. A registration body can ask for information tailored to its statutory process without setting a universal norm for literary acknowledgments. A human author can disclose assistance and remain accountable for a published work.
The responsibility boundary is the line that holds these distinctions together. It marks the point at which a human person or institution remains answerable for what is presented to the public: the claims, expression, source use, publication representation, and relevant rights/ethics obligations. The boundary does not erase other contributors. It does not decide legal authorship in every jurisdiction. It prevents a system’s participation from becoming an alibi for human non-responsibility.
3. What scholarly policies can—and cannot—teach
Scholarly publishing policies are useful because they make certain responsibilities explicit. COPE states that authors using AI tools in writing, image/graphic production, or data collection/analysis should disclose how and which tool was used, while authors remain fully responsible for their manuscript.[1] Taylor & Francis likewise says that authors are accountable for originality, validity, integrity, and references; it does not permit AI tools to be listed as authors and requires disclosure of the tool name/version, use, and reason for use in its publication contexts.[3]
These policies do not treat all activity alike. Taylor & Francis permits forms of assistance such as idea exploration, language improvement, search, classification, and coding support under conditions of human oversight, while warning about inaccuracy, attribution gaps, confidentiality, intellectual-property risk, and potential reuse of inputs/outputs on third-party platforms.[3] Its policy also separates author, editor, and reviewer roles: editors and peer reviewers must not upload unpublished manuscripts or related material into generative systems because of confidentiality and proprietary-rights risks.[3]
Sage offers another bounded threshold model. It calls certain grammar, language, or structure suggestions assistive and says those uses do not require disclosure in its policy, while generated content that directly affects methodology, analysis, results, or conclusions must be disclosed on submission.[5] The distinction is publisher-specific and scholarly. Its teaching value is not that every writer should use Sage’s categories. Its value is that disclosure thresholds can turn on the material role of assistance, not merely whether software with an AI label appeared somewhere in a workflow.
| Policy lesson | What can be translated | What must remain bounded |
|---|---|---|
| AI cannot accept human publication responsibility. | A tool should not be treated as an author who can answer for a work. | COPE and publisher policies govern their own research/publishing contexts. |
| Human review and verification matter. | A responsible disclosure should not imply that a system checked its own claims. | The exact validation requirement differs by genre, risk, and publication process. |
| Tool role and reason can be relevant. | Material assistance can be described more usefully than a generic “AI was used.” | A full tool/version/process record may be appropriate for an editor but unnecessary for a reader. |
| Confidentiality shapes acceptable use. | Private, unpublished, sensitive, or third-party material requires special care. | Policies do not make every prompt or note publicly reportable. |
| Disclosure threshold can vary by impact. | Assistance affecting substantive content deserves different attention from minor editorial help. | No single publisher’s threshold automatically governs creative work. |
The translation must therefore be careful. A research article that reports empirical results has responsibilities different from a poem, memoir, commissioned essay, news story, marketing text, or private proposal. There may still be good reasons for disclosure in each case. But context is not an excuse for opacity; it is the condition for meaningful transparency.
4. A constructed four-role workflow
Consider a writer preparing a public essay. The example is constructed. It does not resolve a copyright question, report a real author’s practice, or endorse a particular tool. It separates four kinds of assistance that a one-sentence note would blur.
| Workflow role | Constructed activity | Potential material question | Possible responsibility response |
|---|---|---|---|
| Private ideation | The writer asks for themes, counterarguments, or possible metaphors, then develops an independent outline. | Did the system’s suggestions materially shape a published claim or expression? Did the writer enter private/unpublished material? | Retain private judgment about whether a public note is relevant; protect confidential process. |
| Language revision | The writer uses a system to flag repetition and propose alternative wording, then accepts, rejects, and rewrites suggestions. | Was the role limited to editorial assistance, or did it create substantive text that survived without meaningful human revision? | Verify final language; use publisher policy/context to decide whether disclosure is material. |
| Source discovery | The system proposes possible articles, statutes, or reports; the writer retrieves and checks original sources. | Did the writer verify every source and claim? Did suggestions introduce invented or misattributed material? | Cite and rely on original sources; retain a source-check record where appropriate. |
| First-pass paragraph generation | The system produces a paragraph later substantially revised by the writer. | How much of the final published expression or argument depends on generated text? What does the intended audience need to know? | Review, revise, and accept responsibility; consider a concise material-use note plus a fuller editorial record if required. |
The purpose of the table is not to declare one activity clean and another prohibited. Different publications will set different standards. The point is that the activities have different relationships to final expression, factual claims, rights, confidentiality, and reader expectations. An accountable practice cannot be built from a question as thin as “Was AI used?”
The constructed workflow also demonstrates why human responsibility cannot be outsourced. A source-discovery tool may point to a relevant report, but only a human writer or editor can decide whether the original source was checked and used fairly. A language suggestion may become part of final prose, but the person publishing must decide whether it is accurate, appropriate, and attributable. A first-pass paragraph may be transformed beyond recognition or may remain materially present. A proportionate disclosure decision depends on that difference.
5. The Responsibility Boundary Canvas
The Responsibility Boundary Canvas is this paper’s original teaching instrument. It helps a writer, editor, publisher, or educator map an AI-assisted workflow before choosing what to disclose, document, retain, or protect. It is not a compliance form and does not decide legal ownership.
| Canvas field | Question | Example output | Failure the field prevents |
|---|---|---|---|
| 1. Assistance type | What precisely did the system do? | “Assisted with language alternatives and surfaced possible sources.” | Treating every use as identical. |
| 2. Material effect | Did assistance affect published expression, factual claims, media, research choices, or a core decision? | “A generated first pass informed a section later materially rewritten.” | Disclosure triggered by a tool label rather than effect. |
| 3. Human review and responsibility | Who checked accuracy, sources, interpretation, originality, and final form? | “The author verified cited sources and approved the final text.” | A note that implies system verification or shifts responsibility. |
| 4. Audience and purpose | Who needs what information, and why? | “Editor receives use details; reader receives a concise material-use note.” | One disclosure statement imposed on every audience. |
| 5. Record and retention | What minimal internal record supports accountability? | “Source-verification notes and key workflow decisions retained by author/editor.” | Assuming public disclosure is the only record. |
| 6. Privacy, confidentiality, and rights | What must be withheld, redacted, or specially reviewed? | “Unpublished interview material and private drafts were not entered or are not disclosed.” | Publishing prompts or logs that expose protected process or third-party material. |
The canvas begins with role, not brand. “Used Tool X” is rarely enough because the same system can serve as a spelling assistant, a translator, a search interface, a draft generator, an image maker, or a data-analysis aid. It then asks whether the assistance materially affected what a relevant audience is being asked to trust. Materiality will differ by genre. In a research article, generation affecting methods or conclusions may require detailed disclosure. In a public essay, a reader may need a concise note only when generated or transformed content materially shapes the published work. In an internal editorial process, a publisher may need more information about confidentiality and verification than an eventual reader does.
The canvas also resists a false move: treating disclosure as a way to transfer responsibility. A sentence such as “Some paragraphs were AI-generated” is not a substitute for reviewing facts, checking references, assessing bias, obtaining permissions, or accepting editorial consequences. COPE’s accountability principle and Taylor & Francis’s author-responsibility policy support this narrow claim in their publication settings.[1] [3] The paper extends it as a teaching principle: a disclosure can locate responsibility; it cannot replace it.
6. Disclosure is not documentation
A useful public disclosure is usually brief. A useful internal record may be more detailed. The two should not be confused.
Cwik’s 2026 framework names the distinction directly: disclosure communicates that generative AI was used, while documentation aims to make the workflow traceable, inspectable, and accountable.[6] Its proposed approach distinguishes minimal from extended documentation by task/risk and emphasizes task specificity, proportionality, role specificity, and privacy sensitivity.[6] The source is an open-science framework, not a general creative-writing rule. Yet its distinction answers a growing problem in all publishing contexts: a public audience may need a meaningful signal without needing a writer’s complete private process.
A proportionate practice might operate across three layers.
| Layer | Primary audience | Appropriate content | What it should not become |
|---|---|---|---|
| Public note | Reader, viewer, listener, or purchaser. | Concise statement of material assistance when context makes it relevant; may name a broad role and human responsibility. | A dump of every prompt, draft, or incidental editorial interaction. |
| Editorial record | Editor, publisher, commissioning organization, or institution. | Use type, material effect, verification steps, confidentiality considerations, and relevant policy compliance. | A blanket demand for all private ideation regardless of risk. |
| Protected working record | Author and any necessary accountable collaborator. | Source-check notes, important decisions, redactions, or records needed for later clarification. | A permanent surveillance archive accessible to any interested party. |
The layers can align without being identical. A writer may tell readers that an AI system helped generate preliminary text that was materially revised, while retaining private notes that show how sources were checked. An editor may receive enough information to assess a publisher policy without asking for personal journal entries or unpublished drafts. A journalist may disclose a material use in a location that serves audience understanding while preserving sources and private work.
The distinction is protective as well as practical. Taylor & Francis warns that third-party AI platforms can create confidentiality, data-security, and intellectual-property risks, especially for unpublished manuscripts and review materials.[3] RTDNA emphasizes that AI policy must consider privacy and fundamental rights in addition to accuracy, context, trust, and transparency.[4] A disclosure practice that forces the publication of sensitive prompts, third-party material, protected research data, or private creative drafts can violate the very responsibilities it claims to promote.
7. The reader-trust counterpoint
Journalism offers a useful but non-transferable counterpoint. RTDNA states that AI should not replace human judgment and critical thinking in trusted reporting and encourages newsrooms to develop clear policies concerning newsgathering, editing, and distribution.[4] Its guidance asks practical questions: Can the organization verify accuracy? What is the audience’s interest? How is AI defined for the purpose of policy—generation, distribution, organization, editing, grammar? Where will disclosure appear? Can journalists review AI-influenced content before it reaches the public?[4]
These questions do not convert creative writing into journalism. They show why audience and purpose matter. A disclosure is not made in a vacuum. A report involving public facts, a news image, a personal essay, a literary experiment, a research publication, and a private draft present different expectations, risks, and privacy conditions. The responsible writer or publisher should be able to say why a disclosure is present, what it communicates, and what human review occurred.
RTDNA also provides a useful caution against generic transparency rhetoric. It frames disclosure in relation to public benefit and risk to trust, while recognizing privacy.[4] The implication for this paper is modest: more information is not automatically more truthful. A public note should reveal a material fact that helps an audience understand the work. It should not merely announce a brand name, perform moral purity, or expose a process that is irrelevant to readers and harmful to the writer or others.
8. A separate legal and registration boundary
Public disclosure is not the same as a legal registration application. The U.S. Copyright Office’s 2023 guidance concerns its own copyright-registration process. It treats human authorship as a case-by-case inquiry and instructs applicants to disclose AI-generated content in works submitted for registration, describe the human author’s contributions, and exclude/disclaim more-than-de-minimis AI-generated material under its process.[2]
That guidance is important for one reason here: it demonstrates that different audiences can require different information for different purposes. A registration body may need a description of human and AI-generated components. A publisher may need to evaluate a manuscript under its policy. A reader may need only a concise material-use note. A writer may retain a protected working record. These are related accountability practices, but not interchangeable ones.
This paper offers no legal advice. It does not determine whether a particular work is copyrightable, whether a contribution is more than de minimis, which jurisdiction applies, or what an individual should place in an application. The Copyright Office source itself is U.S.-specific and requires case-by-case analysis.[2] The Responsibility Boundary Canvas is a publishing and teaching instrument, not a legal test.
9. A bounded action rule
A responsible disclosure practice can be summarized in six steps:
- Name the assistance role. Do not start with a tool brand or an abstract claim that AI was involved.
- Assess material effect. Ask whether assistance affected expression, claims, sources, media, research choices, or a core editorial decision in a way a relevant audience should understand.
- Locate human responsibility. Identify who verified, revised, approved, and can answer for the work.
- Choose the audience. Separate reader-facing, editorial, institutional, and registration needs rather than forcing one statement to serve all.
- Retain proportionate records. Keep enough information to support accountability where needed; do not mistake a public note for full documentation.
- Protect privacy, confidentiality, and rights. Do not disclose prompts, drafts, sources, personal information, or unpublished material merely to satisfy a generic transparency demand.
The guiding principle is neither “disclose everything” nor “disclose nothing.” It is contextual clarity with human responsibility. A writer should not be able to hide material system participation that affects what a relevant reader or publisher is being asked to trust. A publisher should not demand exhaustive creative-process disclosure when a concise public note and a protected internal record would better preserve truth, privacy, and accountability.
The purpose of disclosure is not to declare a machine an author. It is to make the human responsibility boundary legible.
Limitations
This is a conceptual teaching manuscript, not a legal opinion, copyright-registration guide, publisher policy, empirical study of reader response, or universal AI disclosure standard. The four-role workflow is constructed and does not describe a particular author or tool. COPE, Taylor & Francis, and Sage sources address scholarly/publisher settings; their norms are translated cautiously rather than imposed on all creative or professional writing. RTDNA’s guidance is journalism-specific and is used for audience, human-review, trust, and privacy questions rather than as a general literary rule. The U.S. Copyright Office source is a U.S. registration-policy document and requires case-by-case analysis; it does not determine the legal status of any specific work here. Cwik’s source is an abstract/structured summary of a scholarly documentation framework and is used for a bounded disclosure/documentation distinction, not as proof that one record format is always sufficient. The Responsibility Boundary Canvas is an original instrument intended to support deliberation; it cannot settle an individual’s legal, contractual, ethical, or employment obligations.
References
[1] Committee on Publication Ethics, “Authorship and AI tools,” 2024.
[2] U.S. Copyright Office, “Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence,” 88 FR 16190, 2023.
[3] Taylor & Francis, “AI Policy.”
[4] Radio Television Digital News Association, “Use of Artificial Intelligence in Journalism.”
[5] Sage, “Artificial intelligence policy.”
[6] J. C. Cwik, “Disclosure is not documentation: an open science framework for documenting generative AI use in scholarly research and publication workflows,” 2026.