AI for Creators · B guide

When Should You Disclose AI-Generated Content?

A useful AI disclosure answers the reader’s practical question: what was made or materially changed, why it matters here, and what human review still occurred.

Updated September 14, 2026 · Editorial source review

An editorial review sheet showing an AI-use decision tree, three sample labels, and a release checklist.

Disclose AI use when it could reasonably change how a reader interprets a claim, image, endorsement, authorship, or decision. Use a short, specific label at the relevant content, then let the applicable law, platform rules, and editorial policy each do their separate jobs.

Disclosure is an interpretation tool, not a confession

Readers do not need a production diary for every spelling suggestion or background cleanup. They do need context when AI involvement could alter what they think they are seeing, hearing, buying, or relying on. A generated product image may look like a real photograph. A cloned voice may sound like a real speaker. A drafted recommendation may sound like a personal endorsement. In those situations, a label helps a reader interpret the work before acting on it. The practical question is not “Did any software use AI?” It is “Would a reasonable reader make a different judgment if this origin or alteration were hidden?” If the answer might be yes, plan a disclosure at the point of encounter.

  • Focus on material meaning, not tool novelty.
  • Place context where the affected media or claim appears.
  • Keep the label readable without opening a separate policy page.

Use a three-branch decision tree

First, identify the output: text, image, audio, video, recommendation, or endorsement-like claim. Second, ask whether AI created, simulated, or materially changed a feature that a reader could mistake for a real person, event, product result, authorship, or independent opinion. If no, retain an internal production note only if your workflow needs one. If yes, ask whether a law, regulator, contract, or platform has a more specific rule; follow that rule and do not replace it with a vaguer house label. Finally, apply the publication’s editorial standard: disclose enough to prevent a misleading impression, name the review that actually occurred, and hold the item if facts, permissions, or policy ownership are unresolved. This is a workflow, not legal advice for every jurisdiction.

  • No material interpretive change: document internally if useful.
  • Possible material change: check applicable external rules and label the item.
  • Unclear ownership, facts, or permission: hold for the responsible reviewer.

Write a label that says what happened

A generic “AI-assisted” badge can conceal more than it clarifies. Match the wording to the reader’s decision. Sample label one, for an illustration: “AI-generated illustration. It is conceptual and does not depict a documented event.” Sample label two, for edited audio: “This audio was synthetically recreated from an approved script; it is not a recording of the named person.” Sample label three, for editorial text: “AI assisted with an initial draft; an editor reviewed sources, claims, and final wording.” Each label is short, specific, and tied to a real process. Do not claim human review, consent, or verification unless it happened and the responsible owner can identify what was reviewed.

  • State whether media is generated, simulated, or materially altered.
  • Say what the asset is not when that prevents a likely misunderstanding.
  • Name only completed review steps, not aspirational controls.

Keep law, platform rules, and editorial policy separate

These three layers can overlap, but they are not interchangeable. A legal or regulatory obligation depends on the applicable jurisdiction, audience, claim, and facts; obtain appropriate advice when the decision requires it. A platform may require labels, metadata, or special handling for altered or synthetic media even where no general rule appears to demand the same wording. An editorial policy can choose a clearer standard for reader trust, such as labeling realistic conceptual images or recording AI use on sensitive explainers. The FTC’s endorsement materials and AI-claims guidance are useful primary references for U.S.-facing marketing questions, but they do not replace a fact-specific legal review. Treat the strictest applicable, verified requirement as the floor, then write the editorial label in plain language.

  • Law: fact- and jurisdiction-dependent obligations.
  • Platform: product-specific publishing rules and controls.
  • Editorial: the publication’s reader-facing honesty standard.

Make the disclosure part of the release path

Add five fields to the asset or draft record: intended use, AI role, human role, source or permission status, and disclosure decision. A review can then ask a visible question: “What would a reader reasonably assume without this label?” For example, before publishing a generated event image, confirm it is marked as illustrative, that the caption does not imply attendance, and that any associated factual claim has an independent source. Before publishing a voice recreation, confirm the script, speaker identity, permission basis, platform rules, and delivery channel have been reviewed by the responsible owner. The record should support an editorial decision, not become a substitute for it.

  • Attach the label decision to the current asset version.
  • Check captions, headlines, and thumbnails for contradictory implications.
  • Escalate permissions, identity, or advertising questions before release.

Recognize the common failure modes

A disclosure fails when it is too far from the media, too vague to change interpretation, or contradicted by the surrounding headline. “Enhanced with AI” beside a realistic fictional product demo may still leave a viewer believing the feature exists. A hidden footer cannot correct an endorsement-style video that implies a real customer spoke. Metadata alone is also not a complete reader disclosure: it may be stripped, unavailable, or unfamiliar to the audience. Conversely, labeling every minor drafting aid can create noise that hides the material cases. The repair is to identify the misleading inference, then state the minimum accurate fact that corrects it at the moment a reader needs it.

  • Failure: label says nothing material. Fix: name the relevant alteration.
  • Failure: label is distant or hidden. Fix: place it with the media or claim.
  • Failure: surrounding copy overpromises. Fix: revise the copy and the label together.

Run a release checklist before publication

Before release, confirm: the item’s AI role is described accurately; the label is visible in the intended crop, embed, and share context; captions and calls to action do not create a conflicting impression; every factual claim has its own source; any person, voice, likeness, or third-party material has been routed to the responsible owner; applicable platform requirements have been checked for the actual account and format; and the named human review really occurred. The U.S. Copyright Office and C2PA materials can inform provenance and authorship questions, but neither turns an unlabeled or misleading presentation into a trustworthy one. If a required fact is unknown, pause publication rather than inventing a safe-sounding label.

  • Preview the disclosure on the actual publishing surface.
  • Keep the current version and review decision together.
  • Hold unresolved cases for the accountable legal, platform, or editorial owner.

Make corrections as visible as the original impression

A disclosure plan also needs a correction path. If a later source shows that an image was presented with the wrong context, a clip was materially altered, or a label omitted a relevant fact, first stop any scheduled reuse of the asset. Then correct or remove the item where readers are most likely to encounter it, update the caption or article note with the verified change, and retain the original review record. Do not quietly replace a file while leaving a headline, thumbnail, social post, or embedded player that still creates the old impression. The correction should describe the publication’s verified action without speculating about intent or making a legal conclusion. This is especially important for assets that can be copied into newsletters, partner posts, or later compilations. A clear record helps the next editor avoid repeating the same error.

  • Stop scheduled reuse while the material question is reviewed.
  • Correct the context on the surfaces where the claim was made.
  • Preserve the decision record and the reason for the correction.

Sources and update note

Follow the linked official source before a product, price, plan, or policy decision.