EU AI Act's labeling rules go live: mark AI output or risk 3% fines
Article 50 is now in force: chatbots must disclose they're AI, generated media needs machine-readable marks, and fines reach €15M or 3% of turnover.
Copy markdown
What went live August 2
Article 50 of the EU AI Act — its transparency layer — became enforceable on August 2, 2026. It binds anyone whose AI reaches EU users, splitting duties between providers who build the models and deployers who ship the apps.
Your chatbot has to out itself
Any AI that talks directly to people — support bots, agents, voice avatars — must make clear it's a machine at first contact, unless that's already obvious. If you run a customer-facing agent in the EU, that disclosure is now legally required, not a nicety.
Generated media needs a machine-readable mark
Providers of generative systems must mark synthetic audio, images, video, and text in a machine-readable format — C2PA-style provenance or watermarking — that's detectable as AI-made. Narrow carve-outs cover source code, short number/symbol strings, and standard editing assists.
Deepfakes and public-interest text get visible labels
Beyond the hidden marks, deployers must clearly flag deepfakes to viewers at first exposure, and label AI-written text on public-interest topics like politics, health, and justice — unless a human editor takes responsibility for it.
The fine is €15M or 3% of turnover
Non-compliance carries penalties up to €15 million or 3% of worldwide annual revenue, whichever is higher. A missing disclosure banner now sits in the same fine tier as far heavier obligations.
You get until Dec 2 if you shipped early
Generative systems already on the market before August 2 have until December 2, 2026 to meet the machine-readable marking rule, and content both generated and published before August 2 needs no retroactive labeling. New systems get no grace period — wire the marking in now.