Article 50 sets new rules for AI content
AI watermarks are the new cookies
ChatGPT, Gemini, and Claude will now embed cryptographic signatures, invisible to the human eye, into AI-generated texts, images, and videos12. That should make AI detectors much more accurate, because right now their accuracy is close to tarot-card readings.
The European Commission has approved labels reading "AI," "AI Modified," and "AI Generated," in a minimalist black-and-white design3. You will probably start seeing these labels outside Europe too.
An agent has to say what it is, in the first message
AI agents must now disclose that they are machines in the very first message to a person. The AI should clearly state: "You are interacting with an AI system" - the wording proposed by the European Commission in its guidelines4.
When introducing itself, an AI agent must also disclose which person or organisation it represents. Now, when my agent sends an email, it must begin with: "I'm an AI acting on Pavel's behalf."
The fine, and who collects it
In Germany, enforcement falls to the Federal Network Agency, Bundesnetzagentur5, which could fine me 3% of my annual turnover - or up to €15 million if I were a large legal entity6.
My prediction: Pavel Durov will be one of the first people to receive a multimillion-euro fine. He has traditionally been at odds with the European Union, and Telegram's recent Guest AI Bots update is a direct violation of Article 50.
"But if I do not live in the European Union, how can Bundesnetzagentur punish me?" It can ask Apple or Google to remove your application from their app stores, a procedure that is already well established.
Journalists, bloggers, and other people who post images
If you publish AI-generated content on a matter of public interest, you must label it "AI Generated."
The AI Act also applies to people who publish videos and images. If an AI-generated image contains a real person (for example, Donald Trump), or a real place (let's say the Bundestag), it qualifies as a deepfake7 and must carry a label. If you break the rule, the fine can again reach €15 million.
References
- Regulation (EU) 2024/1689 of the European Parliament and of the Council laying down harmonised rules on artificial intelligence (the AI Act), Article 50 - the transparency obligations that apply from 2 August 2026 - https://eur-lex.europa.eu/eli/reg/2024/1689/oj↩
- Article 50, Transparency Obligations for Providers and Deployers of Certain AI Systems - the requirement that synthetic output be marked in a machine-readable format - https://artificialintelligenceact.eu/article/50/↩
- European Commission, Regulatory framework for AI - the Commission's own summary of the obligations and their dates - https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai↩
- Coalition for Content Provenance and Authenticity (C2PA) - the provenance standard the marking is built on - https://c2pa.org/↩
- Bundesnetzagentur - the German market surveillance authority for the AI Act - https://www.bundesnetzagentur.de/↩
- Article 99, Penalties - up to €15 million or 3% of worldwide annual turnover for a breach of Article 50 - https://artificialintelligenceact.eu/article/99/↩
- Article 3, Definitions - point 60 defines a deepfake, which is what brings an image of a real person or place under the labelling rule - https://artificialintelligenceact.eu/article/3/↩