Disclose AI Use: A Practical Decision Guide

Disclose what changes the audience's understanding

Disclose AI use when a governing policy requires it or when the tool materially shaped what the audience is being asked to trust. Consider authorship, generated facts or quotations, synthetic media, sensitive decisions, and reasonable audience expectations. State what the tool did, what people checked, and who remains responsible. Routine spelling assistance may not need an article-level note, but internal records and the stricter applicable rule still matter.

There is no single disclosure sentence that fits a newsroom investigation, a student essay, a client brochure, a novel, a research paper, and an internal meeting summary. Anyone offering one probably also owns a universal charger that fits nothing in the drawer.

This guide is an editorial decision framework, not legal or compliance advice. Laws, contracts, platform rules, academic policies, professional standards, and employer requirements vary and change. Identify the rules that actually govern the work before applying a general norm.

Begin with the governing rule

List the authorities in order:

1. Law or regulation for the relevant jurisdiction and use
2. Contract, client terms, collective agreement, or confidentiality duty
3. Professional, academic, research, or funding rules
4. Publisher, employer, school, or platform policy
5. Project-specific promise to sources, contributors, or the audience
6. Editorial judgment where no stricter rule resolves the question

Follow the strictest rule that applies. A platform may allow a use that a client contract prohibits. A publisher's public policy may permit assistance that an academic assignment forbids. “The tool has a button for it” ranks nowhere on this list.

Save the policy version or link checked, date, decision maker, and result. If the rule is ambiguous and the stakes matter, ask the responsible policy owner or qualified adviser rather than converting uncertainty into a cheerful footnote.

Use a materiality test

Ask whether knowing about the AI involvement would reasonably change how the audience evaluates any of these:

• Who authored or witnessed the material
• Whether words are a quotation, reconstruction, translation, or generated simulation
• Whether an image, voice, video, or event depicts reality
• Whether facts came from reporting and sources or from generated text
• Whether a recommendation or decision received qualified human judgment
• Whether a person's likeness, work, or confidential information was used
• Whether the publication followed a stated method

If yes, disclosure is likely useful even when a generic policy page exists. Put the information near the affected material or in a methodology note a reader can find at the relevant moment.

Materiality is not measured by the number of generated words alone. A tool might produce one synthetic quotation that changes the entire article, or help brainstorm fifty unused headings that change nothing the reader receives.

Distinguish assistance from substitution

Create a simple use record:

The labels are descriptive, not moral grades. “AI-assisted” can conceal anything from a comma suggestion to a rebuilt article. A useful disclosure names the material task.

Avoid saying “written with AI” when the tool only created an outline, and avoid saying “lightly assisted” when entire passages were generated and retained. Precision builds more trust than either alarm or euphemism.

Write a disclosure that answers three questions

A practical statement covers:

1. What did the tool materially do?
2. What did people do to review or transform the output?
3. Who is responsible for the published result?

For example:

An AI tool was used to group the author's interview notes by theme. The author checked every grouping against the recordings, selected and verified the quotations, wrote the article, and approved the final text.

Or:

The illustration was generated from an art-direction brief and then edited by the design team. It is an illustrative scene, not documentary evidence of an event.

These are patterns, not universal wording. Adjust them to the facts and policy. Do not claim “human reviewed” if the review consisted of noticing that the document opened.

Name the product and model only when a rule requires it or the detail helps the audience understand limitations or reproduce the method. Product names, versions, and features change; a durable disclosure can often describe the task while the internal log preserves technical detail.

Put the note where it can do its job

Choose placement by relevance:

• Byline or contributor note: authorship and drafting context
• Methodology box: reporting, analysis, summarization, or verification process
• Caption or adjacent label: synthetic or materially altered media
• Endnote: bounded assistance that matters but does not interrupt the main text
• Policy page: routine organization-wide practice
• Project record: internal accountability, including uses that do not need a public label

A global policy does not automatically explain an unusual use in one article. An adjacent label may be essential when an image could be mistaken for a photograph. Conversely, repeating a vague AI sentence on every page can become wallpaper: technically visible, informationally asleep.

Use plain language and accessible presentation. Do not hide meaningful information behind an unfamiliar icon, color alone, hover-only interaction, or a link called “learn more” with no context.

Handle synthetic media as a distinct case

Generated or materially altered images, audio, and video can affect what viewers believe they are seeing or hearing. Ask:

• Does it depict a real person, event, place, product, or piece of evidence?
• Could a reasonable viewer mistake it for a capture of reality?
• Does it use a person's likeness or voice?
• Is alteration central to criticism, satire, accessibility, reconstruction, or design?
• What nearby label and provenance should remain when the asset is shared?

The C2PA Content Credentials explainer describes a technical method for recording an asset's origin and modification history. It also makes an important distinction: provenance can help establish where an asset came from and what happened to it, but it does not decide whether the depicted claim is factually true.

That means a credential is not a truth badge, and the absence of one is not proof of deception. Preserve available provenance, apply the governing label requirements, and still verify the content and context.

Protect authorship and credit

Do not list a tool as though it accepted responsibility, granted permission, resolved conflicts, or could answer a correction request. Credit the people who performed the accountable work under the publication's rules.

The U.S. Copyright Office's January 2025 summary says its report on AI output copyrightability centers human-authored expressive elements and distinguishes human creative selection or modification from prompting alone. Copyright treatment is fact-specific and jurisdiction-dependent; this source is not a universal authorship rule. It does reinforce an editorial point: record what people actually contributed rather than treating the prompt box as a co-worker with paperwork.

For collaborative work, document who reported, drafted, edited, verified, designed, translated, and approved. If generated material draws on a contributor's work or likeness, address consent, rights, and contractual obligations separately from the disclosure sentence.

Treat facts and citations as a verification issue

Disclosure does not rescue unsupported content. “AI-assisted” is not another way to say “the citations may be imaginary.” Use the claim-ledger workflow to open sources, record supporting passages, recompute numbers, and remove material that cannot be verified.

Likewise, a factually accurate article may still need disclosure because the method, authorship, or synthetic media is material. Accuracy and transparency overlap, but one does not substitute for the other.

When a tool suggests sources, record that internally if relevant. Publicly, cite the sources that support the claims—not the tool that guessed useful search terms.

Document even when public disclosure is unnecessary

An internal AI-use record can include:

• Date, tool, and version when known
• Account or approved environment
• Task and material supplied
• Privacy and permission decision
• Output incorporated into the work
• Human revisions and reviewers
• Claims and sources checked
• Public disclosure decision and governing rule
• Final approver

Scale the record to risk. A spelling suggestion needs less documentation than generated analysis used in a public recommendation. The purpose is to reconstruct material decisions, investigate a problem, and improve the workflow—not to preserve every autocomplete as a tiny historical artifact.

Connect this record to the version and decision log in the AI editing workflow. If a team uses standard prompts, keep the tested versions from the four-pass prompt guide with the process documentation.

Avoid five unhelpful disclosures

“Created using cutting-edge AI.” This is marketing, not process information.

“AI may have been used.” The publisher should know whether material use occurred.

“Human reviewed.” Say what was checked and who approved it.

“No AI was used.” Make this claim only when the workflow can support the scope, including contractors and embedded tools under the applicable definition.

A tool logo with no text. Readers should not need product archaeology to understand an authorship or synthetic-media signal.

Also avoid apologizing for routine assistance when no deception occurred. Disclosure should inform, not stage a morality play around the spell-checker.

Apply a final decision checklist

Before publication, ask:

• Which rule governs this use, and when was it checked?
• What material task did the tool perform?
• Did generated expression, facts, quotations, translation, analysis, or media enter the final work?
• Could omission mislead the audience about authorship, evidence, reality, or review?
• Does the note state meaningful human work and responsibility?
• Is it placed where the affected audience can find it?
• Are facts independently verified and rights handled separately?
• Does the internal record support the public statement?

When in doubt, do not solve the problem with vagueness. Narrow the use, ask the policy owner, seek qualified advice where needed, or write a more specific note. Good disclosure is not the longest note. It is the one that gives the audience the context they need without pretending transparency can replace judgment.