| Rule | Who it is for | What they must do | Starts |
|---|---|---|---|
| 50(1) Tell people it is AI | Companies whose AI talks with people | Tell people they are talking to an AI, unless it is obvious | 2 Aug 2026 |
| 50(2) Mark AI output | Companies whose AI makes text, images, audio or video | Mark the output as AI-made so it can be detected, as far as possible | 2 Aug 2026 for new tools. 2 Dec 2026 for tools already on sale |
| 50(3) Emotion and face systems | Users of emotion or face analysis systems | Tell the people it is used on | 2 Aug 2026 |
| 50(4) Deepfakes | People and businesses who publish deepfake images, audio or video | Say the content is made or changed by AI | 2 Aug 2026 |
| 50(4) AI text on public matters | People and businesses who publish AI text to inform the public on matters of public interest | Say it is AI-made, unless a human editor checked it and takes responsibility | 2 Aug 2026 |
EU AI Act Article 50 checker
textGrain exists because of Article 50 of the EU AI Act. It asks AI companies to mark what their AI makes, and asks some publishers to label it. Answer a few questions to see which rules apply to you, and from when.
Article 50 checker
1Which describes you best?
Pick the one closest to how you use AI.
2Does your AI output reach people in the EU?
Users, readers or customers in any EU country.
3Do people chat or talk with your AI directly?
For example a chatbot, voice assistant or AI agent.
4When did your AI tool go live in the EU?
This sets your marking deadline.
5Which AI model does it use?
6Do you publish realistic AI images, audio or video of real people, places or events?
7Do you publish AI-written text to inform people about public matters?
For example news, politics, health or public safety.
8Does a person check or edit it, and take responsibility for it?
For example an editor at your company.
Your result
General information based on the AI Act and the European Commission’s guidance from October 2026. Not legal advice.
Companies that make AI tools must add hidden marks to AI output. This applies from 2 August 2026 for new tools and from 2 December 2026 for tools already on sale. Publishers must label deepfakes, and AI text about matters of public interest unless a human editor checks it. Personal use is not covered.
Article 50 in short
Where textGrain fits in
OpenAI makes ChatGPT, so Article 50(2) says it must mark what ChatGPT writes. textGrain is how OpenAI does this for text. It is turned on for ChatGPT and Codex users in the EU.
If you build your own app on OpenAI’s API, the same rule may apply to you. OpenAI lets API customers turn textGrain on for some models. It is off by default, so check your settings and ask a lawyer if it covers your product.
If you publish AI content, the hidden watermark is not your label. The European Commission says publishers cannot use the AI company’s hidden mark to meet their own duty to tell readers.
Not legal advice. This page sums up the AI Act and the European Commission’s draft guidance as of October 2026. The guidance may still change.
Key dates
- 1 Aug 2024
The AI Act becomes law.
- 27 Jul 2026
The Digital Omnibus on AI becomes law. It moves the high-risk rules to December 2027, but keeps the Article 50 date. It gives a short extra period for marking only.
- 2 Aug 2026
The Article 50 rules start.
- 5 Oct 2026
OpenAI announces textGrain for ChatGPT and Codex in the EU, and as an option in the API everywhere.
- 2 Dec 2026
AI tools that were on sale before 2 August 2026 must now mark their output.
Questions people ask
Do I have to label text I write with ChatGPT?
Not under the AI Act if it is only for personal use. If you publish AI text for work to inform the public on matters of public interest, you have to say so. The exception is when a human editor checks it and someone takes responsibility for it.
Does marketing text need an AI label?
The label rule covers text that informs the public on matters of public interest. Normal marketing and product text is usually not that. But advertising rules and platform rules may still ask you to say it.
What counts as a human editor checking the text?
A real review or edit by a person, with a person or company taking responsibility for what is published. A quick look before you press publish is probably not enough. Keep a record of who checked what.
How big are the fines?
Up to €15 million or 3% of yearly worldwide sales, whichever is higher. For small and medium businesses, the lower of the two applies.
Is a hidden watermark enough?
For AI companies, a strong watermark or metadata is the kind of mark the law has in mind. For publishers it is not enough. Their label has to be visible to people.