Europe’s chatbot disclosure now starts with the first exchange

A customer-service box can look human long before it says anything useful. It may have a friendly name, type in short bursts and apologise for a delay. Europe’s newest AI transparency rule makes the opening moment more important: when an AI system is directly conducting the conversation, the person should know that from the start rather than discovering it halfway through the exchange.

Article 50 of the EU AI Act began applying on 2 August 2026. Its best-known provisions cover deepfakes and the marking of synthetic content, but the first paragraph deals with something more ordinary. Providers must design AI systems intended to interact directly with natural people so those people are informed that they are dealing with AI, unless that is already obvious in the circumstances.

The Commission’s final guidance turns that broad principle into a useful product boundary. It does not say that every automated feature needs the same notice. It asks whether there is a real, direct conversation between an AI system and a person.

The Commission’s FAQ sets out four cumulative criteria for this part of the rule. The product must meet the Act’s definition of an AI system. It must be designed for a genuine two-way exchange, rather than simply collecting information or issuing a fixed automated response. The AI itself must communicate directly with the person instead of passing its output through a human intermediary. The other participant must be a natural person, whether a consumer, employee or professional user.

That creates a clearer distinction between a chatbot and the software sitting behind it. A conversational assistant that receives a question and generates a reply is the obvious case. An AI suggestion shown privately to a human support worker is different for this specific notice duty because the employee, not the AI, speaks to the customer. A risk score passed between systems, a recommendation engine operating in the background or machine-to-machine traffic also falls outside this particular direct-interaction test.

Outside this test does not mean outside the AI Act or every other law. A background system may be covered by data protection, consumer, employment, sectoral or high-risk AI rules. The point is narrower: Article 50’s conversation notice is triggered by the form of the interaction, not merely by the presence of AI somewhere in the service.

Article 50 says the information must be clear and distinguishable, meet applicable accessibility requirements and arrive no later than the first interaction. The Commission’s FAQ sharpens that to the start of the first interaction.

For a product team, that moves disclosure out of the long privacy page and into the place where the exchange begins. The rule does not prescribe one universal badge, phrase or screen for interactive AI. It does require the information to do its job before a person has already calibrated the conversation as human.

Context still matters because the law contains an exception where the AI interaction is obvious to a reasonably well-informed, observant and circumspect person. The guidance says that exception should be interpreted restrictively. A product called an AI assistant may make the situation plain. A human name, portrait, conversational delay or customer-service styling can make it less plain. The test is not whether the product team knows how the system works. It is whether an ordinary person can tell what they are interacting with in that setting.

This is also why a notice is not the same as a warning. The rule is about identity and informed trust, not a claim that the system is dangerous, inaccurate or incapable. A calm line at the opening can establish the nature of the exchange without turning every chatbot into a red-alert screen.

The timing can be confusing because Article 50 contains several different duties. Providers of generative AI systems must also add machine-readable marks to certain synthetic outputs. The Commission says a limited transition until 2 December 2026 applies to that marking duty for relevant systems placed on the market before 2 August.

The FAQ is explicit that the limited grace period concerns Article 50(2), the marking and detection obligation. It is not presented as a general postponement for chatbot identity notices. The direct-interaction rule began applying on 2 August along with the other transparency provisions.

The roles differ too. The provider is responsible for designing a directly interactive AI system so people are informed. Deployers carry separate duties in areas such as emotion recognition, biometric categorisation, deepfakes and certain unreviewed public-interest text. A single service can therefore have more than one transparency question, but they should not be collapsed into one generic “AI used here” footer.

The practical value of the rule is not the number of times the letters AI appear on screen. It is whether the interface preserves an honest boundary between person and machine.

That boundary can be tested at the opening. Who is speaking? Is the response generated as part of a two-way exchange? Is a human reading and choosing what to send? Is the AI role already unmistakable to an ordinary user? Can the notice be perceived with assistive technology and understood before the conversation develops?

Those questions do not determine every legal case, and the Commission’s guidelines are guidance rather than a substitute for the Act or an authority’s decision. They do make the reader-facing consequence unusually concrete. In Europe, the first useful line from a chatbot may now be the one that tells you it is a chatbot.