How to write a disclosure that satisfies the rules
How to write an AI disclosure that satisfies the EU rules: worked examples, placement, tone, and the review that beats any wording.
Part of our guide to the EU AI disclosure rules.
This guide is not legal advice. It is a practical guide to writing the actual sentences, with examples you can adapt. For what the law requires and of whom, start with our main guide.
The standard you're writing to
The Act asks for one thing, said plainly in Article 50(5): the disclosure reaches people "in a clear and distinguishable manner at the latest at the time of the first interaction or exposure". Clear, findable, and there when the content is. Everything in this guide is that sentence applied.
Notice what the standard is not. It is not a legal formula, a mandated wording, or a badge scheme. Nobody can sell you the officially correct sentence, because there is none. What there is, is a test any visitor could apply: did the site tell me, where I could see it, before I had consumed the thing it was telling me about?
What good sounds like
A good disclosure is short, specific, human, and placed where the content is. Some worked examples, free to adapt.
For an article a model helped draft and a person reviewed: "Parts of this article were drafted with AI assistance and reviewed before publication by a named member of our team." If the named person is happy to be named in public, better still: named disclosures read as confidence, anonymous ones read as small print.
For a page that is substantially AI-generated: "This page was generated with AI and is checked by our team." At the top, not the bottom, because the reader should know before they have read it, not after.
For an AI-generated image: "This image was generated with AI." As a caption or an adjacent line. Six words. For a page of them, one clearly placed note with the images can serve, provided a visitor meets it alongside the images rather than three scrolls later.
For machine-translated pages: "This page was translated with AI assistance and reviewed." One sentence per language, in that language, which is the part people forget: a disclosure in English on a German page discloses very little to the German reader.
For a chatbot: the greeting says what it is. "You're chatting with our assistant, which is automated." No human name, no stock photo of a person who does not work there.
What lazy sounds like
The genre has three classics, and all three fail the same test.
The footer mumble: "Some content on this site may be AI-generated." Site-wide, non-specific, and three scrolls from anything it describes. It tells the visitor nothing about the page in front of them, which means it discloses nothing. This is the cookie-banner school of compliance: words in the shape of transparency.
The buried policy: a diligent page at /ai-policy, unlinked from anywhere a human goes. A disclosure the visitor must excavate is not clear, and it is certainly not encountered at first exposure. It is a confession filed where nobody will read it.
The weasel: "We may use various tools in producing content." May. Various. Tools. A sentence engineered to be technically consistent with anything, which is exactly why it discloses nothing. If the honest version feels too blunt to publish, the discomfort is information about the content, not the disclosure.
Placement, quickly
Articles: top or foot of the piece, and the top is braver and better if the piece is substantially generated. Images: caption or adjacent line. Whole pages: before the content, not after. Product listings generated at scale: a clear note on the listing template, once, does the work everywhere. The single test, always: would an ordinary visitor meet the disclosure by the time they meet the content? If yes, placement is solved. If it took you a moment to decide, it is not.
The tone question
Write the disclosure in your own voice, because a disclosure that sounds like your site reads as honesty, and one that sounds like a legal department reads as damage control. Compare "Parts of this guide were drafted with AI assistance and reviewed by our team before publication" with "Certain content herein may have been produced utilising automated generative technologies." Same fact. The first sounds like a company telling you something; the second sounds like a company hoping you would stop reading, and you did.
There is also a quiet commercial truth here that surprises people: done plainly, disclosure reads as confidence. The businesses that say "a model drafted this, a person checked it, here is who" look like they know what they are doing, because they do. The ones who hide it look like they have something to hide, right up until someone checks.
The part that outranks the wording
Here is the thing this whole guide has been walking towards. For text, the Act's most generous provision is not about disclosure wording at all: AI-generated text that "has undergone a process of human review or editorial control", with a named person holding "editorial responsibility", needs no disclosure whatsoever. The best disclosure, for reviewed content, is a review you can prove.
Which reframes the writing job. The sentences in this guide are for the content you choose to label. For the content you would rather stand behind, the work is a real review, a real name, and a real record, and the wording takes care of itself, because there is none to write. A signed, recorded human review beats the most elegant disclosure sentence ever drafted, in the same way that being on time beats a beautifully worded apology.
Images, as ever, are the exception: no review lifts the image duty, so their six words stay regardless. Our images guide has that story.
Doing it at any scale
One site, one afternoon: walk the pages, write the sentences, decide what gets reviewed instead, and record who reviewed it. Ten sites or a hundred: that walk is a scan, the sentences want drafting in bulk, and the reviews need a system with a record, or they will quietly stop happening by March.
Both of those are what Swornmark does: findings with disclosure wording already drafted in your register, a named sign-off flow, and a certificate that proves the review happened. The scan is free, and if your disclosures are already good, it will be the shortest report we ever send.
Questions people actually ask
What is an example of a good AI disclosure statement? For a reviewed article: "Parts of this article were drafted with AI assistance and reviewed before publication." For a substantially generated page: "This page was generated with AI and is checked by our team", at the top. For an image: "This image was generated with AI", as a caption. Short, specific, in your own voice, and placed with the thing it describes.
Where on the page should an AI disclosure go? Wherever an ordinary visitor meets it by the time they meet the content. In practice: top of a substantially generated article, caption or adjacent line for an image, the greeting for a chatbot, the listing template for products generated at scale. If you have to think about whether a spot qualifies, it does not.
Is "some content may be AI-generated" in our footer enough? No. It is site-wide, non-specific and three scrolls from anything it describes, so it tells the visitor nothing about the page in front of them, which means it discloses nothing. The Act asks for clear and distinguishable at first exposure, and the footer mumble is neither.
Do we need a separate AI policy page? Have one if you like, but it does not satisfy the duty. A disclosure the visitor must excavate from /ai-policy is not encountered at first exposure; it is a confession filed where nobody will read it. The policy page can explain your approach. The disclosure still lives with the content.
Do we have to write the disclosure in every language our site publishes in? Yes, in that language. An English disclosure on a German page discloses very little to the German reader. One sentence per language, on the pages that need it, which is less work than it sounds.
Can we skip the disclosure entirely if a person reviews the text? For text, yes, that is the Act's own trade: a real review process, a named person holding editorial responsibility, and a record that proves it happened. For images there is no such lift; their six words stay regardless of who looked at them.
This guide is not legal advice. It was drafted with AI assistance and reviewed before publication. Swornmark holds editorial responsibility for it. You will notice that sentence keeps appearing at the foot of these guides. That is what a disclosure looks like.
More guides
Does my website need AI disclosure?
A decision-tree guide to whether the EU's AI disclosure rules reach your website, with the questions people actually ask.
AI images and the label rule: images are the ones that bite
Why AI-generated images are the strictest corner of the EU's disclosure rules, and what a label a visitor can see actually means.
What Article 50 actually says
A clause-by-clause walk through Article 50 of the EU AI Act, quoting the text and translating it for people who run websites.
The agency problem: client sites and whose duty it is
When an agency builds and runs a client's website, whose AI disclosure duty is it? The practical answer for agency owners.
Regulators, country by country: who would actually come asking
Which national authorities enforce the EU's AI disclosure rules, country by country, and how enforcement actually reaches a business.
The state of AI disclosure in Europe, August 2026
We scanned 494 ordinary business websites in 17 EU countries, 26 days after Article 50 took effect. Under 1% label their AI content. The numbers, by country and sector, with the method.
The EU's AI disclosure rules, explained for people who run websites