If you build or run AI that reaches people in the EU, you now have to tell them when they're talking to a machine, and more.
Starting August 2, 2026, Article 50 of the EU AI Act applies to any business that develops or deploys AI whose outputs are meant for use in the EU.
That brings concrete rules: systems that interact with people must disclose at first contact that a machine is replying. Companies that deploy emotion recognition or biometric categorisation must notify individuals when those systems are operating, whether in real time or applied later. AI-generated text and deepfakes about public-interest topics must be labelled, with a narrow exception where content has undergone human editorial oversight.
The European Commission's AI Office can now demand technical documentation, open investigations, and impose penalties, including fines up to 3 percent of global annual turnover for certain breaches.
Why this matters: these are legal obligations that reach providers and deployers worldwide if their outputs target EU users. Product, legal and trust teams will need to change how features, content pipelines and disclosures are built.
Think of it as a transparency label for AI. Before a conversational system answers, it must first say it's a machine. If a system reads emotions or classifies biometric traits, people should be told it happened. Labelled deepfakes make public-interest content easier to spot.
What's next: regulators must show they will use these powers, and companies must show they can operationalise the rules. The real test is whether enforcement targets large model providers or focuses on how AI is deployed. Either way, product design is about to change.