When should publishers disclose AI assistance?

By SimonSeptember 16, 2026

A useful AI disclosure names the task, the checks that actually happened and the person accountable. Build a policy around what readers could otherwise misunderstand.

Translucent paper layers with indigo tabs and a blank note card held by a dark paperweight.

AI-generated conceptual illustration: visible layers and a source note represent explaining AI assistance and editorial checks.

"AI assisted" could mean a spelling suggestion, a translated interview, or an article generated from documents nobody checked. The label leaves the reader to guess.

A useful disclosure says what the tool did, what someone checked and who takes responsibility for the published version. Put it near the material it explains. A policy in the footer won't help a reader understand that an apparent photograph depicts something that never happened.

Set the policy by task, with stricter checks for factual drafting and realistic synthetic media. A byline can't tell readers which checks took place.

Start with what readers could misunderstand

Reuters Institute's 2024 research on public attitudes towards AI in journalism found different reactions to different uses. In qualitative work with 45 digital news users in Mexico, the UK and the US, participants were more comfortable with behind-the-scenes assistance than with generating new content. Realistic synthetic images and video raised particular concern.

Those findings concern news audiences in a particular period. They give you questions to ask your own audience, not proof that readers of tutorials, entertainment or product content feel the same way.

Reuters' own AI policy offers one threshold: it says it discloses when it relies primarily or solely on generative AI to produce news content, and provides context about that use. That is Reuters' rule, not a universal one. Google's guidance on helpful, reliable content separately recommends considering disclosures where readers would reasonably wonder how content was created, particularly when automation substantially generates it.

Your policy can be more specific. Would knowing the AI's role change how a reasonable reader interprets the work as original reporting, an exact quotation, independent testing, a translation or an illustration? If it would, explain that role where the reader needs to see it.

Decide what each task requires

Adapt this worksheet to your subject matter and applicable obligations. The checks in the final column need a real person to do them before you can claim they're complete.

AI task: Spelling or formatting assistance

Suggested reader-facing treatment: Cover routine use in the publication policy unless meaning changed

Evidence required before claiming review: Editor confirms the final meaning and wording

AI task: Interview transcription

Suggested reader-facing treatment: Usually cover in the policy; add a specific note if publishing the transcript as an artifact

Evidence required before claiming review: Quotes checked against the recording, including names and uncertainty

AI task: Translation

Suggested reader-facing treatment: Label the translated version and identify the source language

Evidence required before claiming review: A competent reviewer compares meaning, idiom and material quotations with the original

AI task: Summary of existing reporting

Suggested reader-facing treatment: Label an AI-generated summary near the summary; link to the full reporting

Evidence required before claiming review: Reviewer checks every summary claim and material omission against the source

AI task: Substantial factual drafting

Suggested reader-facing treatment: Use an article-level note identifying the AI's role

Evidence required before claiming review: Accountable reviewer checks claims against sources, rather than merely reading for fluency

AI task: Generated diagram or conceptual illustration

Suggested reader-facing treatment: Caption the visual accurately and identify AI generation where used

Evidence required before claiming review: Reviewer checks labels, relationships and whether it could be mistaken for evidence

AI task: Realistic synthetic image, voice or video

Suggested reader-facing treatment: Require explicit editorial approval and conspicuous labeling; reject uses likely to mislead

Evidence required before claiming review: Document purpose, rights checks and the risk of false attribution

Don't classify the work only by the first tool someone opened. A transcription tool that rewrites a speaker's answer into cleaner prose is doing more than transcription. A translation needs another review if an editor later asks a model to make it more persuasive. Record what happened to the material throughout the process.

A disclosure decision tree separating routine assistance from changes to meaning, authorship or perceived evidence — Suggested disclosure policy: escalate when AI changes what readers think they are reading or seeing. Legal obligations need a separate check.

Say what "human reviewed" involved

Decide what the reviewer must do and keep a short record. A name beside the article should identify someone who has accepted responsibility for that version, but the name alone doesn't explain the review.

For a factual article, record the reviewed version, reviewer, date and any unresolved claims. For an interview, identify the recording used to check quotations. For a translated guide, identify someone who can assess both languages and understands the subject matter.

If nobody performed the check, "editorial oversight" is a misleading way to describe the gap. Hold the material, remove the unsupported claim, or explain the limitation when publication is otherwise justified. An AI-generated second opinion can help find mistakes. It cannot accept responsibility, approve its own work on someone's behalf or replace human sign-off.

The approver also needs authority to stop publication and time to inspect the sources. If the workflow demands a signature without allowing that time, fix the workflow before writing a disclosure that suggests a thorough review.

Give readers a specific note

The following are templates, not descriptions of work performed for this publication. Replace placeholders only with verified facts. If you can't support a claim about review, don't include it.

For substantial drafting:

AI was used to prepare an initial draft from the sources linked in this article. [Reviewer name and role] checked the factual claims against those sources and approved [version/date]. Report errors through [corrections route].

For a translated version:

This is an AI-assisted translation from [language]. [Reviewer] compared it with the original for meaning and terminology. The original is available at [link].

For an illustration:

AI-generated conceptual illustration. It does not depict a real customer, event or product interface.

Keep the note proportionate to the use. A caption can explain a visual without sending readers to another page. Substantial drafting usually needs a short note on the article itself, with a link to the wider policy for details. If a newsletter carries an AI-generated summary on its own, include the relevant disclosure there too. The original page's note won't travel with the summary unless you make that happen.

An annotated disclosure note identifying the AI task, completed checks, accountable reviewer and corrections route — Template anatomy. The reviewer and completed-check fields must be supported by the publication record before use.

Check rights and confidential inputs separately

Disclosing AI use doesn't establish permission to upload an interview, reuse a photograph, imitate a voice or reproduce someone else's material.

Before using the tool, check what inputs it will receive and review the relevant permissions, confidentiality commitments, contracts and tool terms. Record that decision separately from the disclosure decision. Uncertain copyright, privacy or likeness questions need qualified advice for the applicable jurisdiction; this worksheet cannot resolve them.

An accurate "AI-generated" caption can still sit beneath an image with unresolved rights or an editorial reason not to publish it. Keep those approvals separate so the caption doesn't become a substitute for permission.

Your policy should also distinguish using AI to produce work from allowing outside services to retrieve it. The crawler-controls guide explains the technical checks for the latter. They are different decisions.

Make someone responsible for corrections

Tell readers where to report a problem and who will respond. Reuters' AI statement, for example, includes a route for concerns about mistakes or misrepresentations due to AI.

Keep enough internal version history to find the affected article, image and any newsletter reuse. If a generated summary distorts a source, check everywhere that summary went. Correcting the web page won't correct an email already sent.

Revisit the policy when the work changes: a new language, a new synthetic-media format, or a tool used for factual drafting instead of copy assistance. One useful check is to ask a reader to explain a sample disclosure back to you. If they can't tell what was generated and what was checked, revise the note.

The disclosure should describe the work as it happened, including its limits. Claim human accountability only when a real person has accepted it for the version readers will see.