- If you publish AI images, audio or video that look real (deepfakes), you must disclose it visibly. A hidden watermark from the tool is not enough.
- AI-written text needs a label when it informs the public on matters of public interest, unless a human exercised genuine editorial oversight.
- People must be told when they are talking to a chatbot, unless it is obvious.
- It applies to anyone deploying AI in the EU or reaching EU audiences. Platforms are adding their own labels too.
Checked 28 October 2026. This is a marketer’s summary of Article 50 of the EU AI Act and the Commission’s guidelines and FAQ, as reported by law firms. It is not legal advice; definitions such as “deepfake” and “public interest” are interpreted case by case. Ask a lawyer for anything regulated or political.
01The four duties
Marketers are mostly deployers: you use an AI tool, so the fourth duty is yours. The marking duty sits with the tool’s provider, with a grace period to 2 December 2026 for systems already on the market before 2 August.12
02What a valid label looks like
New season, new colours. Shop the look →
Animation. The Commission’s guidelines expect a disclosure people can actually see.
| Content | Good | Not enough |
|---|---|---|
| Image | Label next to or on the image | A line in a footer or terms page |
| Video | On the opening screen, or clearly visible during playback | Only in the description |
| Audio | Spoken or shown disclosure at the start | None, or only metadata |
| Text (public interest) | Below the headline or at the top of the piece | Buried in an about page |
The Commission has published optional icons and a voluntary Code of Practice. Using them is not required; the disclosure itself is.3
Do I need to label this?
Answer four questions about one piece of content. The decider follows Article 50 as the guidelines describe it, and tells you why.
Runs in your browser; nothing is saved. A guide, not legal advice. Platform rules (Meta, TikTok, YouTube) may require labels even where the law does not.
03Marketing cases
AI model wearing your product
A photorealistic person who does not exist, in a real-looking setting, is a deepfake under the Act. Disclose it on the image.
Cloned founder voice-over
AI audio imitating a real person must be disclosed, even with their consent.
Illustrated or stylised visuals
Content nobody would mistake for a real photo or recording is generally outside the deepfake rule. Platforms may still ask for a label.
AI-drafted blog post
Commercial content is not “public interest”. Health or news-style content is, unless a human editor took real responsibility for it.
Customer-service chatbot
Say it is a bot at the start of the conversation, unless it is obvious from context.
The tool’s watermark
Do not strip metadata or watermarks when exporting. The provider’s marking is part of the system.
Platforms are moving in the same direction: Meta, TikTok and YouTube already apply or require AI labels on realistic synthetic media. Google’s helpful content guidance now treats fake AI authors as deception.
04A labelling routine
Tag AI assets at creation
Add an “AI” field to your asset library. If it was generated or substantially altered, mark it there first.
Build the label into templates
A small, consistent badge for images and an opening-screen card for video. Decide the wording once.
Keep provider marks intact
Export settings that preserve metadata; no re-saving through tools that strip it.
Review public-interest text
Anything about health, money, safety or news gets a named human editor, or a label.
Disclosure is not an apology for using AI. It is what keeps the audience’s trust when they find out anyway.
Common questions
Do I have to label AI-generated content in the EU?
Since 2 August 2026, Article 50 of the EU AI Act requires deployers to disclose deepfakes (AI images, audio or video that appear real) and to label AI-generated text published to inform the public on matters of public interest, unless there was genuine human editorial oversight. Providers must also mark AI outputs in a machine-readable way.
Is a watermark enough to comply with Article 50?
No, for deepfakes. The machine-readable marking is the provider’s duty; the deployer who publishes the content must also disclose it in a way people can clearly see, near the content itself.
Does old AI content need to be labelled?
No. The Commission says content generated before 2 August 2026 does not need to be labelled retroactively.
Do AI-written product descriptions need a label?
Generally no. The text labelling duty applies to text that informs the public on matters of public interest, such as news or health, not ordinary commercial copy. Platform rules may still apply.
Method & sources
Checked October 28, 2026. Based on the European Commission’s FAQ and guidelines on Article 50 and law-firm analyses of them. Placement guidance is drawn from commentary on the guidelines and may be applied differently by national authorities.
- European Commission, “Transparency obligations under Article 50 of the AI Act” (FAQ). digital-strategy.ec.europa.eu
- Orrick, “EU AI Act: transparency obligations for AI-generated content (Article 50)”. orrick.com
- Reed Smith, “Transparency obligations for AI-generated content: the Code of Practice adequacy decision and the final Commission guidelines”, and European Commission, “Signing the Code of Practice on transparency of AI-generated content”. reedsmith.com · digital-strategy.ec.europa.eu
- WebsiteInit, “EU AI Act and AI-generated content: what must be labelled from 2 August 2026?”. websiteinit.com