The EU AI Act's Transparency Rules Just Went Live: What Article 50 Actually Requires
The EU's AI Act transparency rules took effect August 2, requiring chatbots to disclose they're AI and deepfakes to be labeled. Here's what Article 50 actually covers, who it applies to, and how it compares to California's SB 942.
Contents5
What Article 50 requires, in plain terms
Article 50 covers four separate disclosure obligations, and it's worth treating them separately because they apply to different systems:
- **AI interaction disclosure.** Any chatbot or interactive AI system has to tell a person they're talking to a machine, not a human, and it has to say so by the time of first interaction, not buried in a terms-of-service page three clicks deep.
- **Synthetic content marking.** Content generated or altered by AI needs a machine-readable mark so platforms and tools can detect it automatically. The Commission published a standard set of icons for this.
- **Emotion recognition and biometric categorization disclosure.** If a system detects emotions or sorts people into biometric categories, it has to say so to the people it's analyzing.
- **Deepfake labeling.** Images, audio, or video that have been edited or generated to look real must be labeled as manipulated or AI-made.
None of these obligations require the underlying AI system itself to be safe, accurate, or fair. They're disclosure rules. The Act's other chapters cover risk classification and higher-stakes obligations; this piece is specifically about the transparency layer that kicked in today.
Who actually has to comply, and what breaking the rule costs
The obligations apply to providers and deployers of AI systems operating in the EU, regardless of where the company is headquartered, which is the same extraterritorial logic that made GDPR a global compliance headache rather than a European one. Penalties for non-compliance reach €15 million or 3% of a company's worldwide annual turnover, whichever is higher. For a large AI vendor, 3% of global revenue is a bigger number than the flat fine, so this isn't a rounding-error penalty for anyone with real scale.
Enforcement sits with the AI Office at the Commission level plus each member state's own market surveillance authority, so a company operating across several EU countries could in theory hear from more than one regulator.
How this lines up against California's SB 942
The EU isn't alone in moving on AI transparency this year. California's [AI Transparency Act (SB 942) also took effect on August 2](https://questloops.com/blog/california-s-ai-transparency-act-sb-942-is-now-in-effect-what-changes-on-august-2), and the two laws overlap in intent but differ in mechanics.
| EU AI Act (Article 50) | California SB 942 | |
|---|---|---|
| Trigger | Any AI system operating in the EU | Providers with 1M+ monthly CA users |
| Chatbot disclosure | Required at first interaction | Not the primary focus |
| Content provenance | Machine-readable marks, standard icons | C2PA-compatible provenance in images/video/audio |
| Detection tool | Not mandated | Free public detection tool required |
| Penalty | €15M or 3% global turnover | Set separately under California law |
The practical effect for a company shipping AI image or voice tools to both markets: you now need two separate provenance and disclosure implementations, because the technical requirements (icon-based marking versus C2PA metadata) don't map onto each other one for one. Tools that generate synthetic voice or video, like [ElevenLabs Voice Engine v3](https://questloops.com/blog/elevenlabs-voice-engine-v3-explained-what-changed-and-how-to-try-it-for-free) or image models such as [FLUX 3](https://questloops.com/blog/flux-3-is-here-what-actually-works-in-black-forest-labs-new-model-right-now), sit squarely in scope for both rules if their output reaches EU or California users.
What this means if you're just a user
If you're chatting with a support bot on a European site starting today, it has to tell you it's a bot. If you see AI-generated video or images from major platforms, look for the disclosure icon the Commission published. This won't stop bad actors from ignoring the law entirely, but it does mean the large, compliance-conscious platforms, the ones with EU revenue to protect, will start labeling things they didn't bother labeling last week.
The honest caveat: enforcement on day one rarely looks like raids and fines. It looks like legal teams scrambling to add a disclosure banner before regulators start actually sampling products. Expect the real test of this law to show up in enforcement actions over the next six to twelve months, not today.
A related rule kicked in the same day
Article 50's disclosure rules aren't the only part of the AI Act that activated on August 2, 2026. The Commission's enforcement powers over general-purpose AI model providers, [under Article 53](https://questloops.com/blog/the-eu-ai-act-s-gpai-enforcement-powers-are-live-what-article-53-requires-from-ai-labs), also switched on the same day. That rule targets the labs building the underlying models rather than the companies deploying chatbots and deepfake tools, and it's worth reading if you're choosing between AI providers for a product headed into the EU.
One concrete example of a lab responding to this exact transparency layer: Anthropic's own synthetic content marking took effect a few days late and only for its newest models. [Claude's watermarks are live, but not on the models most people are actually using](https://questloops.com/blog/claude-watermarks-are-live-but-not-on-the-models-you-re-using) breaks down exactly which Claude models are covered so far and which aren't.