Article 50 of the AI Act starts applying on 2 August 2026 — and unlike most of the Regulation, it reaches systems that are already live.
Published in partnership with Dude Hack.
As of today, any AI system that talks directly to people in the EU has to make sure those people know they are talking to a machine.
That is the headline duty in Article 50 of the AI Act, and the European Commission’s own guidance on the provision is unambiguous about the date: it applies as from 2 August 2026, with providers and deployers expected to comply from that point.
What actually kicks in
Four things, aimed at different parties.
Providers of systems that interact directly with people — chatbots, AI agents, avatars — must design them so users are informed they are dealing with AI. Providers of generative systems must mark synthetic audio, image, video and text in a machine-readable format so it can be detected as AI-generated.
Deployers carry the other half. They must tell people when emotion recognition or biometric categorisation systems are in use, and clearly label deepfakes and AI-generated text published to inform the public on matters of public interest.
The part organisations keep getting wrong
This applies to systems already in production. It is not a rule for things built after today.
The only concession is narrow: for AI systems placed on the market before 2 August 2026, the marking and detection obligation under Article 50(2) does not bite until 2 December 2026. Everything else applies now. Content generated before today does not need retroactive labelling, though the Commission encourages it.
Enforcement sits mainly with national market surveillance authorities, with the AI Office holding a limited role. Fines can reach €15 million or 3% of total worldwide turnover, with proportionality available for SMEs and small mid-caps.
Geography does not help either. Providers established outside the EU are in scope where the output of their system is used in the EU.
Two exceptions that will decide most cases
The chatbot duty does not apply where it is obvious a person is dealing with AI. The Commission says this should be read restrictively, since it removes transparency, and sets the test against an average person who is reasonably well-informed and observant.
The labelling duty for published text falls away where content has had human review or editorial control, with someone holding legal responsibility for publication. Crucially, the Commission specifies that superficial checks — it gives spell-checking as the example — do not count.
Where this lands in consumer products
The obvious cases are customer-service bots and AI writing tools. The less obvious ones may generate more argument.
Dating apps have been shipping AI features into conversations for over a year. And on subscription platforms, chatbots answering paid subscriber messages are an established practice: when two OnlyFans subscribers sued over agency-run inboxes in April 2025, Vice noted that some of those services employ people while others run bots.
That places a fairly unglamorous corner of the internet squarely inside Article 50. The Commission’s definition of a deployer covers anyone using an AI system professionally — and explicitly includes individuals who derive economic benefit on a regular basis. A creator or agency running a chatbot in a paid inbox is a deployer, not a private user.
What to watch next
The marking obligation is the harder engineering problem, and it has a four-month runway before the December deadline closes the gap for existing systems. Watermarking synthetic text robustly remains unsolved in a way that watermarking images is not.
The Commission has approved a voluntary Code of Practice on transparency of AI-generated content as a route to demonstrating compliance. Firms that skip it can still comply by other means, but should expect more requests for information about how.
For anything conversational shipping into the EU, though, today’s requirement is simple enough to implement this afternoon: say so, before the first exchange, in a way a person can actually notice.
