AI disclosure statements: what to say, when you need one, and copy-paste examples

"Made with AI" is becoming a required line, not a courtesy. The awkward part is that most disclosure today is a sentence you type and everyone agrees to believe. That works until it does not, and the platforms know it, which is why the rules are tightening and why the more serious answer is to make the claim checkable. Start with what the statement is and when you owe one, then close the honor-system gap.

What an AI disclosure statement is

An AI disclosure statement is a short, plain-language notice that artificial intelligence was used to create or substantially change a piece of content. It names the role AI played, an image was AI-generated, a chapter was drafted by a model, a voice was synthesized, so the audience is not misled about what they are looking at. In its weakest form it is one sentence in a caption or an author's note. In its strongest form it is a machine-readable record attached to the file. Both are called disclosure; only one can be verified, which is the whole tension in this piece.

When you actually need one

There is no single rule. Obligations come from the platform you publish on and, increasingly, from law. The consistent pattern across policies is that they target AI-generated or realistically altered content, and they explicitly exempt light assistance. Three of the most common surfaces:

• YouTube. Creators must disclose realistic altered or synthetic content, meaning material a viewer could mistake for a real person, place, or event. The label appears in the description, or on the video itself for sensitive topics. Clearly unrealistic, animated, or production-assistance uses do not require it.
• Amazon KDP. Authors must declare AI-generated text, images, or translations through a mandatory checkbox when publishing or updating a book. AI-assisted work, brainstorming, editing, refining your own writing, does not need to be declared. Enforcement tightened through 2025, and undisclosed AI content can get titles blocked.
• Steam. Game developers must disclose player-facing AI content on the store page through the Steamworks content survey, separated into pre-generated assets that ship with the game and live-generated content produced during play. As of a January 2026 update, behind-the-scenes efficiency tools like coding assistants are explicitly out of scope.

Beyond platforms, the EU AI Act's Article 50 moves the whole field toward marking synthetic content in a machine-readable way, which is a stronger bar than a human-readable sentence. If your audience is global, design for that direction now.

Copy-paste AI disclosure examples

These are starting points, not legal text. Keep them specific and truthful: name what was AI-made and what was human, because a vague or inaccurate disclosure is worse than none. Replace the bracketed parts.

For an article or blog post

"Parts of this article were generated with [model or tool]. A human author [wrote / edited / fact-checked and approved] the final text. Any images marked AI were created with [tool]."

For a book (KDP-style)

"This book contains AI-generated [text / images / translation] created with [tool]. The content was reviewed and edited by [author name] prior to publication."

For an image or video

"[AI-generated with [tool]] / [Contains AI-altered footage]. No real person depicted is making a real statement. Edited by [name / studio]."

For software or a game

"This [app / game] uses AI to generate [assets / dialogue / art] with [tool]. [Pre-generated assets ship with the product. / Content is generated live during use, with safeguards to prevent harmful output.]"

A statement is a promise, not a receipt

Here is the gap every honor-system disclosure has: the sentence and the content are not bound together. Copy an AI image away from its caption and the disclosure is gone. Screenshot a video and the label does not travel. Edit a file and nothing in it records that AI was ever involved. The claim lives beside the asset, not inside it, so it is trivially separated from the thing it describes.

That is the same weakness a hidden watermark has from the other direction, and you can watch it happen in our interactive lab: a mark that reads as present one moment collapses after a single edit, silently. A written disclosure fails even more simply, because it was never attached in the first place. Transparency you cannot verify is transparency someone else can quietly strip.

How to make a disclosure verifiable

The fix is to bind the disclosure to the asset in a way that survives handling and resists tampering. That is what C2PA Content Credentials do: a cryptographically signed manifest, carried with the file, that records what tool produced the content and how it was edited. Instead of a sentence a viewer has to trust, you get a signed record a viewer can check, one that reads as valid, invalid, or absent rather than simply being taken at your word. The written statement is still worth making for humans skimming a caption. The Content Credential is what makes it stand up when someone actually looks.

Put simply: write the disclosure for people, and sign the provenance for verification. One is the promise. The other is the receipt.

Frequently asked questions