How to Write an AI Assistance Disclosure That Readers Can Use

A useful AI disclosure states what the tool contributed, what a person checked, who accepts responsibility, and which rule governs the work.

Begin With the Rule That Actually Governs the Work

There is no single disclosure sentence that applies to every article, report, campaign, manuscript, classroom, client, platform, country, or copyright filing. Requirements can come from a publisher, employer, school, professional body, grant, contract, client brief, platform rule, or applicable law. A disclosure that satisfies one venue may be incomplete for another. Before drafting the note, identify the controlling rule, the intended audience, and the point in the workflow when disclosure is required.

For example, the International Committee of Medical Journal Editors guidance for authors says journals should require authors to disclose at submission whether AI-assisted technologies were used and to describe how they were used in the cover letter and submitted work where applicable. That is influential guidance for medical publishing, not a universal rule for every kind of writing. Treat it as a clear example of a field-specific policy, then read the instructions for the venue you are actually using.

Map the Material Contribution Before You Write the Note

Start with a private use record. List the tool or service, version or access date when relevant, the people involved, the files or data provided, and each task the system performed. Separate brainstorming from research, retrieval from synthesis, outlining from drafting, translation from editing, and image generation from image selection. Record what entered the final work and what was rejected. The question is not how many prompts were typed; it is what changed in the published artifact because the system was used.

A material contribution affects content, expression, evidence, analysis, or a reader's interpretation. Generating a paragraph, summarizing interviews, proposing a conclusion, translating a claim, calculating a result, or creating a published illustration is usually more consequential than correcting a repeated space. The threshold is still set by the governing policy. Some venues require disclosure of any generative use, while others distinguish routine assistive features from generation. Do not invent your own exception when the rule is explicit.

• Ideation: questions, angles, keywords, counterarguments, or examples proposed by a model.
• Structure: an outline, sequence, heading set, or content plan that shaped the final work.
• Expression: generated or substantially rewritten text, translation, images, audio, or code.
• Evidence work: extraction, classification, synthesis, calculations, citations, or source suggestions.
• Review: grammar, readability, consistency, accessibility, or issue spotting.

Keep Assistance Separate From Authorship and Accountability

A tool can contribute output without becoming an accountable author. The ICMJE's broader principles for AI in publishing say AI tools should not be listed as authors, humans remain responsible for accuracy and appropriate attribution, and users should be transparent about the tool and purpose. The same page warns that confidential submitted manuscripts should not be uploaded where confidentiality cannot be assured without permission.

A reader-facing note should therefore name the responsible human role, not pretend the system approved its own work. “AI reviewed this article” leaves unanswered who checked the output and who can correct it. Better language names the editor, author, analyst, or organization that selected the material, verified defined elements, approved publication, and remains accountable. If nobody can truthfully accept those responsibilities, a polished disclosure does not make the workflow safe.

Replace “Human Reviewed” With Specific Checks

“Human reviewed” sounds reassuring but has no stable meaning. It could describe a quick skim or a line-by-line comparison with primary sources. State the scope. Did a named role verify facts, replay quotations, reproduce calculations, inspect licensing, compare translations, test code, check alt text, or approve claims within a particular domain? Did that reviewer have the source access and expertise needed for the task?

The voluntary, cross-sectoral NIST Generative AI Profile treats clearly defined human oversight roles, source and citation verification, content provenance, privacy, intellectual property, and record retention as parts of risk management. A publication note does not need to reproduce an internal control system. It should, however, be supported by one. Keep evidence outside the chat: source locators, review status, calculations, permissions, decisions, and the final approved wording.

Protect Privacy, Confidentiality, and Rights Before Discussing Transparency

Disclosure happens after a more basic question: was the material authorized for the tool at all? Do not upload confidential manuscripts, personal data, client records, trade secrets, embargoed findings, licensed databases, or private interviews merely because you plan to mention AI later. Check the service terms, organizational controls, retention settings, contract, consent, and approved-use policy first. Minimize inputs and use stable placeholders when the work can be completed without identities.

A public note should not expose the sensitive material it is meant to protect. “Customer records were anonymized before an approved tool generated a first-pass category list” may be useful if true and permitted. Publishing raw prompts that contain those records would defeat the control. Internally, preserve enough information for audit and correction, but apply access controls and retention rules. Transparency is not permission, and disclosure cannot retroactively authorize an improper upload.

Do Not Turn Copyright Registration Guidance Into a Universal Labeling Rule

Copyright questions require special precision. The U.S. Copyright Office's registration guidance for works containing AI-generated material addresses information applicants provide when seeking U.S. copyright registration. It says applicants should disclose included AI-generated content and briefly explain the human contribution; it also describes how to identify human-authored material and exclude more-than-minimal AI-generated material from a claim. That is guidance for a particular government process, not a general command that every AI-assisted web page carry the same public label.

In its 2025 summary of the Copyright and Artificial Intelligence report on copyrightability, the Office said generative-AI output may be protected only where a human author determined sufficient expressive elements, such as through perceptible human-authored material or creative arrangement or modification, and not through prompts alone. It also said using AI as an assistive tool, or including AI-generated material within a larger human-generated work, does not by itself bar copyrightability.

Those distinctions may affect a registration strategy and the records a creator keeps, but they do not resolve every ownership, licensing, contract, or jurisdictional question. This article is an editorial workflow, not legal advice. If registration, ownership, permissions, employment terms, or potential infringement matters to the project, preserve the creation record and ask a qualified professional about the specific facts.

Write for a Decision the Reader Actually Has to Make

Ask what the audience needs to judge. A journal editor may need the tool, purpose, and manuscript section. A client may need to know whether its data entered an external system and who approved the output. A reader of a product comparison may need to know that AI helped summarize vendor materials but a person tested the products and verified prices on a stated date. A viewer of an illustrative image may need to know it is synthetic and not documentary evidence.

Include details that change that judgment. Usually that means the material task, the human checks, the accountable role, and a policy-relevant limitation. Tool name and version can matter when the governing policy asks for them, when output is meant to be reproducible, or when capabilities differ materially. Token counts, every discarded prompt, and vague praise of “advanced AI” usually add noise rather than transparency.

Place the Disclosure Where It Can Do Its Job

Placement follows purpose. A submission declaration belongs in the venue's required field and may also belong in a cover letter or methods section. A public editorial note should sit close enough to the article, image, methodology, or credits that a reader can find it before relying on the affected material. A client disclosure may belong in the brief, statement of work, delivery note, and approval record. Hiding a material disclosure in an unrelated privacy page makes it formally present but practically useless.

Timing matters too. Tell a client or collaborator before using a tool when approval, confidentiality, or data handling is at stake. Do not wait until delivery to reveal a prohibited workflow. At publication, make the final note match the final artifact rather than the original plan. If the team intended to use AI only for headings but later retained generated paragraphs or an image, update the record and the disclosure before approval.

Keep a Versioned Internal Record

A short public note cannot carry the entire provenance trail. Keep an internal record with the artifact version, tool and access date, task description, approved input class, output retained, source packet, reviewer, completed checks, permissions, disclosure text, placement, and approval time. Link it to the content identifier or release rather than leaving it inside a personal chat history.

When the work changes, decide whether the note must change. A sentence-level edit may not matter; replacing a human-written analysis with generated synthesis might. Record the reason. If a correction reveals that the AI contribution or verification scope was described inaccurately, update the public note where the venue allows, preserve the prior version, and follow the publisher's correction process. A disclosure is part of the publication record, not decoration added once and forgotten.

Use a Four-Part Disclosure Structure

A strong note can often fit into four parts: tool and task, material contribution, human verification, and accountability or limitation. Add the exact tool, version, date, placement, or data-handling statement when policy or context makes it useful. Keep the verbs concrete: proposed, drafted, translated, classified, generated, selected, rewrote, verified, tested, approved.

Sentence-level assistance: “A generative AI tool proposed headings and sentence-level alternatives. The author selected and rewrote the final language, checked every factual claim against the cited sources, and approved the published version.”

Substantive first draft: “An approved generative AI service produced portions of an initial explainer from a human-prepared brief and source packet. The editor reorganized and rewrote the draft, verified quotations and factual claims against the linked primary sources, and accepts responsibility for the final article.”

Illustrative image: “The header illustration was generated from a human-directed concept and selected and reviewed by the editorial team. It is illustrative rather than documentary evidence.”

These are patterns, not safe-harbor language. Adapt them to the actual work and the governing rule. If a policy requires a named service and version, include them. If no personal or confidential data was provided and that assurance matters to the audience, say so only after verifying it. Never use a template to claim checks that did not happen.

A Worked Example: From Vague Badge to Useful Record

Imagine a team publishes a 1,500-word guide based on five government sources. A model suggested the outline, drafted two early sections, and proposed the hero image. The editor discarded one section, rewrote the other, generated and selected an illustrative image, verified every final claim against the five sources, and approved the post. No customer data or unpublished source material entered the tool.

The note “Made with AI” is too broad. It does not distinguish brainstorming from retained text, identify the review, or explain the image. “Human reviewed” is too vague. A better note is: “Generative AI was used to propose the outline, draft part of an early version, and create the header illustration. The editor selected and substantially rewrote the retained text, verified every final factual claim against the linked government sources, reviewed the image as illustrative rather than evidentiary, and approved the published version. No customer or unpublished source data was provided to the tool.”

The internal record should be more detailed: tool and access date, approved brief, the two drafted sections, retained language, image provenance, source checklist, editor, approval timestamp, and final note. If a client contract required prior approval or prohibited generated imagery, the team would need to follow that rule before use; the stronger sentence would not cure the breach.

AI Assistance Disclosure Checklist

• The controlling publisher, employer, school, client, contract, platform, and legal requirements were checked.
• The tool, version or access date when relevant, users, tasks, inputs, and retained outputs are recorded internally.
• The note distinguishes ideation, retrieval, drafting, translation, analysis, editing, and image generation.
• Material AI contributions are described with concrete verbs rather than a generic badge.
• The responsible human role and the exact verification performed are named.
• Sources, quotations, calculations, permissions, licensing, and originality were checked as the work requires.
• Confidential, personal, proprietary, licensed, and unpublished inputs were handled under approved controls.
• The disclosure appears in every required submission field and near the affected public content when useful.
• The public note matches the final artifact, not merely the original plan.
• A versioned record supports correction, audit, and later questions without exposing sensitive prompts.

What a Disclosure Cannot Do

A disclosure cannot make an invented citation true, turn an unauthorized upload into authorized processing, obtain a missing image license, create consent, satisfy an undisclosed contractual term, or transfer accountability to a model. It also cannot prove that a work is accurate merely because a person looked at it. Verification must be appropriate to the claim and performed against evidence the reviewer can inspect.

Nor should the note become a performance of transparency that obscures the human decisions. “Powered by responsible AI” is a marketing claim unless the organization can define and support it. A useful disclosure is narrower and more modest. It says what happened in this artifact, what was checked, what remains limited, and who stands behind the result.

Give AI a Bounded Role and Keep the Human Decision Visible

AI can help turn a use record into plain language, compare a draft note with a policy checklist, or flag missing fields. It should not decide whether its own contribution was material, whether a confidential upload was permitted, whether a copyright claim is valid, or whether the team complied with a contract. Those judgments belong to people with the necessary authority, evidence, and expertise.

The best disclosure is usually short because the underlying workflow is clear. Define the tool's role before use, preserve the record, verify the work, identify the accountable person, and write the note for the audience's real decision. Transparency becomes useful when it rests on controls rather than trying to substitute for them.