EU AI Act transparency rules go live Sunday, dragging every chatbot in Europe into scope

Article 50 disclosure obligations apply from 2 August. High-risk deadlines have already slipped to 2027.

On Sunday, 2 August 2026, Article 50 of the EU AI Act becomes enforceable, and every chatbot addressing a European user falls into disclosure scope. The rule is narrow in language and vast in reach: providers of AI systems designed to interact with individuals must tell those individuals they’re interacting with an AI system, per Norton Rose Fulbright’s read of the enforceability calendar. The obligation attaches to whoever applies the brand or trademark, not merely the upstream model provider, which routes liability squarely at the deployer layer where the industry has spent two years insisting it doesn’t sit.

Article 50 also demands that AI-generated content carry a machine-readable mark. Systems already on the EU market before Sunday get a grace period to 2 December 2026, an accommodation Latham & Watkins traces to the May 7 political agreement on the AI Act Omnibus that entered into force in July. Deployers pick up their own layer: disclose deep fakes, and inform individuals when emotion recognition or biometric categorisation is running.

Penalties are the part that concentrates minds. Latham puts the exposure at up to €35 million or 7% of worldwide turnover for Article 50 violations, with civil liability on top. The European Commission published implementing guidelines on 20 July and opened a Code of Practice for sign-on.

The interesting move is what Brussels didn’t enforce. Annex III high-risk system obligations, originally due Sunday, have slid 16 months to 2 December 2027 under the Omnibus deal. AI embedded in regulated products (lifts, toys) now waits until 2 August 2028. The politically fraught tier gets more runway; the transparency tier, cheap to comply with and popular with voters, ships on time.

Beijing read the same room differently. On 15 July, China’s Implementation Opinions on Intelligent Agents and its Interim Measures for Anthropomorphic AI Interaction Services took effect, the first regulatory framework anywhere to treat AI agents as a category distinct from generative AI. Rimon Law’s July brief flags it as a structural bet, not a stylistic one. Two regulatory regimes, three weeks apart, are now competing to define what an AI system is legally required to admit about itself.