Should people disclose when they used AI at work, school, or in creative projects?

The rule in practice

Think of disclosure as an aid to informed reliance. A reader, reviewer, client, assessor, or collaborator may need to know whether they are evaluating your judgment and original expression, a process, a factual claim, a model-produced result, or a deliverable whose rights and risks need extra review.

Four questions make that practical:

1. Is there a binding instruction? A policy, contract, brief, submission form, syllabus, accessibility accommodation, regulator, professional rule, or platform term is controlling. Follow its specified form, tool restrictions, consent requirements, and recordkeeping. If it prohibits the use, do not use the tool and then try to cure the breach with a disclosure.
2. Did AI materially affect authorship or substance? Material means it generated or substantially rewrote a meaningful portion of the final text, code, image, audio, video, design, analysis, or translation, or it selected, ranked, or transformed content in a way that shaped the outcome. An outline that you discarded is usually different from a generated report that you lightly edited.
3. Did AI affect evidence, decisions, or risk? Treat use as material when it found sources, summarized confidential records, extracted data, performed a calculation, recommended a course of action, generated executable code, or affected health, safety, legal, financial, hiring, educational, or security outcomes. Disclosure should be paired with a check of the underlying work.
4. Would a reasonable recipient expect human-only work or need to know? An examiner assessing a student's skills, a client buying bespoke authorship, a journal reader evaluating methods, or an audience seeing a realistic image may reasonably care even if the AI contribution is small by word count.

This is a rule of relevance, not a purity test. It allows people to use useful tools without pretending that every autocomplete event is authorship. It also prevents a vague "AI was used" label from concealing that a system drafted the key argument, chose the data treatment, generated the pull request, or created the published image.

When disclosure is required, prudent, optional, or distracting

The boundary is contextual. Elsevier's current journal policy, for example, distinguishes basic grammar, spelling, and punctuation checks from substantive changes to sentence structure or organization, which should be disclosed. It also asks authors to state the tool, purpose, and extent of oversight. That is a useful model for proportionality, even outside publishing. Elsevier Generative AI policies for journals

Match the disclosure to the kind of AI use

Brainstorming and outlining

Using a chatbot to produce questions, counterarguments, names, or possible structures is often optional to disclose when you independently develop the final work and no policy says otherwise. It becomes prudent when the prompt meaningfully set the thesis, argument, or creative concept that the recipient assumes came from you alone. It is required in assessments and commissioned work whenever the relevant rules make ideation part of the work being evaluated or bought.

Do not mistake plausible suggestions for sources. If the system suggests a fact, quotation, statistic, case, paper, or citation, locate and read the original source before relying on it. A disclosure such as "AI assisted brainstorming" does not repair fabricated evidence or a missing citation.

Editing, proofreading, and translation

Limited spelling, grammar, punctuation, formatting, and accessibility support are commonly treated as low-impact assistance. A rewrite that changes structure, voice, argument, or tone is different because it may change authorship and could mask the person's actual communication skill. Translation also needs care. Machine translation can be an appropriate tool for an internal draft or routine content, but disclose it when the exact wording has legal, clinical, contractual, literary, cultural, or assessment significance. Have a qualified human check consequential translations.

If AI assisted a person with a disability, avoid treating that assistance as suspicious by default. Follow the applicable accommodation and course or workplace process. Disclosure, where needed, should describe the allowed support without revealing unnecessary personal medical information.

Research, analysis, and recommendations

Disclose AI use when it did more than help you phrase a search. Examples include generating a literature review, coding qualitative data, extracting facts from files, ranking candidates, making a forecast, producing a calculation, or recommending an action. State the inputs or dataset at a meaningful level, the task it performed, important human checks, and known limitations. The recipient needs enough information to evaluate reliability, reproducibility, bias, and responsibility.

The National Institute of Standards and Technology's Generative AI Profile identifies governance, pre-deployment testing, content provenance, and incident disclosure as relevant considerations. It also notes that generative AI can call for additional human review, tracking, documentation, and management oversight. NIST AI RMF Generative AI Profile For a high-stakes analysis, preserve the actual source data, method, prompts or workflow version when appropriate, validation results, and the person who signed off. A public disclosure can remain short while the audit record is more detailed.

Code and technical deliverables

Treat accepted AI-generated code as code you authored for accountability purposes. You must understand the change well enough to maintain it, test it, and defend it in review. Disclose in the pull request, contribution notes, or client handoff when AI generated a material feature, a nontrivial refactor, security-sensitive logic, tests that might shape review, or a large portion of the change. In a repository with an explicit policy, follow that policy even for smaller uses.

Human review is especially important around authorization, authentication, cryptography, payments, personal data, infrastructure changes, production commands, and safety controls. GitHub documents that AI-generated code can be inaccurate or introduce security risks and says outputs should be reviewed and tested before merging, particularly for critical or sensitive applications. GitHub Copilot Agents responsible-use documentation Review code provenance and licenses too. Code-reference features can help surface matches with public repositories, but GitHub notes limits: altered suggestions, private code, code outside GitHub, and newly changed public code may not be covered by a particular check. GitHub Copilot code referencing

Generated images, audio, video, and other media

Disclosure is usually prudent for published AI-generated or materially AI-altered media, and often required on platforms, in news, advertising, research, or professional submissions. It is particularly important where the work could be mistaken for a real person, event, place, testimony, product, or recording. Caption the image, credit it in production notes, or use the publication's label rather than hiding a vague notice in fine print.

For creative projects, disclosure has a second job: it tells collaborators, commissioners, publishers, and audiences what rights and production claims they are being asked to accept. Check contracts, client briefs, performer or likeness permissions, stock licenses, model terms, music and asset licenses, and platform rules before using an input or publishing an output. The World Intellectual Property Organization warns filmmakers that a GenAI service's terms can affect rights in materials uploaded as prompts and that output can create third-party copyright or personality-rights risks. WIPO rights-clearance guide

Substantial authorship

If AI supplied the central expression or most of the final deliverable, describe it plainly. Calling yourself the sole author of a text, illustration, song, treatment, or design may mislead a buyer, editor, examiner, collaborator, or audience if their purpose is to assess your own creative work. Your selection, direction, arrangement, and revision can still be meaningful human contribution, but do not inflate them into a claim that the system merely "assisted" when it created the core output.

Copyright rules vary by country and facts, so this is not legal advice. In the United States, the Copyright Office says that a work using AI may be protected where a human determined sufficient expressive elements, including through human-authored material or creative arrangement or modification, but prompts alone do not establish enough human control. The Office's registration guidance also requires applicants to disclose AI-generated material included in a submitted work. U.S. Copyright Office Part 2 report U.S. Copyright Office registration guidance announcement Disclosing use does not settle ownership, licensing, infringement, labor-credit, or contractual questions.

The settings that override personal preference

Work for an employer or client

The controlling question is normally not whether colleagues approve of AI. It is whether the organization permits this tool for this information and task. Check the acceptable-use policy, information-security rules, approved-vendor list, data processing agreement, client statement of work, confidentiality agreement, intellectual-property terms, and any client instructions. An employer may allow a private enterprise tool but ban public consumer tools, code agents with external access, or uploading sensitive documents.

Do not paste trade secrets, source code, unpublished strategy, personal data, client records, credentials, security details, or regulated information into a system unless your organization has approved that exact data flow. Removing a name may not be enough to anonymize a record. If you discovered an issue after an upload, follow the organization's incident process promptly. A later disclosure to the client does not necessarily remove the confidentiality or contract problem.

For a commissioned deliverable, ask one direct question before beginning: "May I use this tool for this task, and what disclosure, approval, attribution, and rights assurance do you require?" Put the answer in writing when the work is valuable, confidential, or likely to be reused. If the work represents the client's own voice, expertise, or bespoke creative authorship, material AI generation is more likely to require consent even if the final quality is high.

School and assessment

The course instruction is the first authority. Universities take different approaches. Vanderbilt, for example, states that instructors decide how or whether their classes allow AI and that, absent a course statement, students must disclose generative-AI use. Vanderbilt academic integrity guidance Other institutions prohibit use where the instructor does. This variation is why a student should not infer permission from another course, another school, or a general belief that AI is common.

If the instruction is unclear, ask before using the tool and preserve the answer. Explain the intended use specifically: brainstorming, language editing, debugging, translation, feedback on a draft, generating code, or drafting an answer. An assignment may assess the process of reasoning, writing, coding, or translation, not merely the final artifact. A truthful disclosure cannot substitute for work that the assignment required you personally to do.

Journals, research, and professional work

Scholarly rules are often stricter because authorship carries responsibility for accuracy, originality, methods, citations, and confidentiality. The International Committee of Medical Journal Editors says authors, reviewers, and editors should be transparent about AI use, identify the tool and purpose, and not list AI as an author. It also warns that uploading submitted manuscripts to systems that cannot assure confidentiality may violate the authors' rights. ICMJE recommendations on AI use Follow the specific target journal or society policy because forms, locations, and permitted uses differ.

In professional advice, a human remains responsible for the output. Do not present model output as legal, medical, financial, engineering, or safety analysis without qualified review and the evidence that the professional standard requires. If AI affected a conclusion, document the method and validation for the file, and tell the recipient anything needed to interpret or challenge the advice.

Write specific disclosures

A good notice is short, factual, and proportionate. It answers four questions without pretending that the tool deserves authorship:

1. What tool or class of tool was used?
2. What task did it perform?
3. How much did it affect the final work?
4. What did a human do to review, revise, verify, and take responsibility?

Avoid claims such as "AI-free" unless you can define and support them. Modern writing, design, search, translation, and coding products use a wide range of automated features. Be precise about the relevant use instead: generative drafting, image generation, code generation, source synthesis, translation, or editing.

Example disclosures

Report. Hypothetical setup: an analyst wrote a board report after using a language model to propose a structure and summarize meeting notes, then checked each claim against the notes and source documents. Action: place this note near the methodology or in an appendix: "A generative-AI tool was used to propose an outline and create an initial summary of meeting notes. The author reviewed the source notes, rewrote the report, and verified all factual claims." Takeaway: the disclosure identifies material assistance without assigning the model responsibility or exposing confidential prompts.

Code contribution. Hypothetical setup: a developer used an AI coding assistant to draft a new API validation module, then rewrote portions, added independent tests, ran security checks, and received normal code review. Action: add this in the pull request: "AI assistance was used to draft parts of the validation module. I reviewed and modified the implementation, wrote and checked the tests, ran the project security and test suites, and remain responsible for this change." Takeaway: the review record says what matters to maintainers, especially if the repository expects disclosure.

School assignment. Hypothetical setup: a course permits AI for feedback but not for drafting. The student asks an AI tool for a checklist on clarity, rewrites their own prose, and does not copy generated text. Action: submit the permitted declaration: "I used an AI feedback tool to identify clarity issues in my draft. I wrote the submitted response myself and used no AI-generated text or citations." This explains how the use fits the assignment’s permitted assistance. If the syllabus does not allow this, the student should not use it.

Creative work. Hypothetical setup: an illustrator directs an image model to create preliminary backgrounds, selects one, paints over it extensively, and publishes the result in a commercial campaign. Action: use clear production credits: "Background began as an AI-generated concept image and was substantially repainted and composited by the artist. No depiction of a real person was intended." The credit explains the process; licensing, likeness, and ownership still need separate checks. Those need separate clearance.

Verification and records

Disclosure is about transparency. Verification is about whether the work is safe to rely on. Do both. The more consequential the work, the more evidence you should retain. A sensible private record can include:

• the approved tool, account tier, and relevant version or configuration
• the purpose, date, and responsible human
• the categories of input shared, without copying sensitive data into the record unnecessarily
• what output was used or rejected and where it appears in the deliverable
• source checks, calculations, tests, reviews, security scans, and approval decisions
• any public disclosure, permissions, licenses, or client consent

Do not preserve raw prompts or outputs automatically if they contain personal, confidential, copyrighted, or legally privileged material. Your records policy, retention schedule, and privacy obligations still apply. For safety-critical systems, regulated work, or high-impact decisions, use the organization's established quality, traceability, incident-reporting, and independent-review process rather than relying on a generic AI note.

A compact decision rule

Use this default: binding rule first, material contribution second, recipient need third. If a governing rule applies, follow it. If AI materially shaped the work, evidence, or risk, disclose it specifically and verify the result. If it did not, ask whether disclosure changes a reasonable recipient's understanding. If not, a private record is usually enough.

That approach lets people admit ordinary tool use without shame and avoids hiding meaningful automation behind vague language. It also keeps the central responsibility clear: the human who submits, signs, ships, or publishes the work owns the decision to rely on it.