AI labelling means telling a person, in the interface, that what they are looking at was made by a machine.
On 2 August 2026 that stopped being a courtesy. Article 50(4) of the EU AI Act became enforceable. It reaches any company serving people in the EU, and it binds both the people who build the system and the people who deploy it. Fines go up to 15 million euros or 3% of worldwide turnover, whichever is higher.
Most teams responded by keeping the little sparkle icon next to the button. The Commission looked at that and called it too ambiguous to comply.
A sparkle is not a disclosure. It is a decoration that happens to sit near one.
It failed for a reason that has nothing to do with law. On different products in the same week it has meant magic, beta, premium and new. A symbol that means four things means none.
Five Things Your Interface Has to Do Now
1. Say it in words
What teams ship: an icon on its own.
Why it fails: the Commission published an official EU AI icon set, then said plainly that an icon alone does not make you compliant.
Ship instead: plain words, in the reader's language. "AI-generated." Keep the icon if you like it, but the words carry the meaning.
2. Put it where the content is
What teams ship: a line in the footer, or a tooltip behind a hover.
Why it fails: the standard is "clearly visible". Footer notes, faint icons and flashing labels are all named as non-compliant.
Ship instead: put the label next to the thing it describes, at a size you would use for anything you wanted read.
3. Make it survive leaving your app
What teams ship: a label drawn in the UI and nowhere else.
Why it fails: the label has to stay when content is reshared or downloaded. A badge in your React tree is gone the moment someone saves the image.
Ship instead: mark the file, not the screen, with metadata and watermarking. This one has a grace period: systems already on the market have until 2 December 2026.
4. Announce the chatbot first
What teams ship: a friendly name, a human-looking avatar, a note somewhere in the transcript.
Why it fails: people have to know they are not talking to a human. After they have asked their question is too late.
Ship instead: make the disclosure part of the opening state, not part of the conversation.
5. Let a screen reader find it
What teams ship: a decorative icon with no accessible name.
Why it fails: the labelling has to reach assistive technology. A disclosure only sighted users receive is not a disclosure.
Ship instead: real text in the accessibility tree, announced with the content it belongs to.
What Is Actually Covered
| Covered | Not covered |
|---|---|
| Deepfakes of real people or places | AI-assisted work with real human involvement |
| Chatbots and agents talking to people | Closed business-to-business tools |
| Fully AI-written text on public interest topics | Anything published before 2 August 2026 |
| Emotion recognition and biometric sorting | Text with real editing and a named editor |
Public interest is defined broadly: health, safety, environment, money, politics, science, culture. Your changelog is fine. Your AI-written health explainer is not.
The Exemption Everyone Will Misuse
AI-written text is exempt when a human has reviewed and edited it, and a named person takes editorial responsibility. Both halves matter, and teams will skip the second.
The Commission was specific about what does not count: spellcheck, formatting, cropping, translation. If your publishing flow has no field for who took responsibility, you do not have the exemption. You have a habit.
The Four-Question Audit
- Does the label use words, not only a symbol?
- Is it visible at the content, with no hover or scroll?
- Does it survive a screenshot, a download and a reshare?
- Does a screen reader announce it?
Most teams fail three and four. Three is an engineering ticket. Four takes ten minutes and nobody has filed it.
The honest read is not about compliance. The sparkle was failing users long before it failed regulators. The rule removed the option of not saying it out loud.
Get the PDF