On 2 August 2026 the transparency provisions of the EU AI Act — Article 50 — started applying. The headline requirement is disarmingly simple: chatbots and other interactive AI systems have to tell users they are dealing with AI, not a human. Anyone running a customer-facing conversational system should read that sentence twice, because a good many deployments were built on the opposite instinct.
What actually came into effect
The obligations split into four strands, and only the first is about conversation.
The Commission has published guidelines on the transparency obligations to go with them, and the AI Office has issued a voluntary Code of Practice on Transparency of AI-Generated Content — a recognised route to demonstrating compliance with the marking and detection duties, including a set of icons for labelling generated material. The Commission reports that more than 180 organisations have signed it.
Why a UK business should care
Because the AI Act's reach is defined by where the output lands, not where the company is registered. Providers established outside the EU fall within scope when the output of their system is used in the Union. A London agency running an AI chat assistant that serves customers in Dublin or Berlin is inside the tent. So is anyone whose AI phone line takes calls from EU numbers.
The exposure is real rather than theoretical. Breaches of Article 50 sit in the second penalty tier under Article 99(4)(g): fines up to €15 million, or up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. There is a mitigation for smaller firms — under Article 99(6), SMEs and start-ups face the lower of the two figures rather than the higher — but "lower of two large numbers" is a poor foundation for a compliance strategy.
The bit people get wrong
The disclosure has to be in the conversation. For a chatbot, the information must reach the user before or at the very beginning of the interaction, and the Commission's guidance is explicit that burying a reference in terms and conditions does not discharge the duty. The notification also has to account for children and for people with disabilities — which, in practice, means it needs to be legible to a screen reader and comprehensible to someone who is not reading carefully.
On a voice call this is easier than teams expect, and the sensible pattern is well established: the agent says what it is in its opening line, in the same breath as the greeting, without ceremony. It costs about two seconds. The failure mode is not a compliance breach so much as a trust one — callers who work it out for themselves halfway through a conversation feel duped, and a caller who feels duped is not a customer.
What to do this month
Three concrete things. Check that every conversational surface you operate — website widget, phone line, in-app assistant — states plainly that it is an AI, at the start, in the channel itself. Check that the statement survives the awkward paths: the person who arrives mid-conversation, the caller transferred from elsewhere. And write down where your systems' outputs are used, because the answer to "are we in scope" turns entirely on that.
None of this is onerous for a well-built system. It is only onerous for one that was quietly relying on people not asking.
Related: Do you legally have to tell people it's an AI? · Voice AI and chat widgets: what conversational AI actually does · The ICO's crackdown on fake-human AI calls