Your chatbot has to say it is a chatbot: the EU AI Act rule that landed on 2 August
On 2 August 2026 the transparency obligations of the EU AI Act started to apply. The headline one is short enough to fit in a sentence: if a person is interacting with an AI system, they have to be told. Most websites running a chat bubble have not changed a thing, which is why this is worth twenty minutes of your week.
What Article 50 actually asks for
The obligation is that people are informed they are interacting with an AI system, at the point of interaction. Not in a policy page, not in a footer — at the moment the conversation starts. A parallel obligation covers marking AI-generated content in a machine-readable way; generative systems already on the market have until 2 December 2026 to comply with that part.
The obvious exception is thinner than it looks
Disclosure is not required where it is obvious from the point of view of a reasonably well-informed, observant and circumspect person that they are dealing with AI. Published guidance cautions against leaning on that. A chat panel in the corner of a shop that answers instantly at three in the morning is obvious to you, because you built it. It is not obvious to a sixty-year-old buying garden furniture.
It binds more people than you would expect
Providers have to design the disclosure in. Organisations deploying somebody else's system cannot assume it has been handled for them — the duty has to sit in their own compliance and governance framework. And the reach is territorial by user, not by company address: serving EU users is enough, wherever you are established.
What good compliance actually looks like
- Say it in the first message the visitor sees, not in a link they will not click.
- Make it survive language switching — a disclosure that only exists in English is not a disclosure for a Portuguese visitor.
- Keep it visible on every surface the agent runs on: widget, hosted page, WhatsApp, Telegram.
- Keep a record of what was disclosed and when, which mostly means keeping your transcripts.
- Mark AI-generated content where you publish it, ahead of the December date.
Where SuperCognit already helps
The welcome message and behavior settings are edited visually, so adding a line is a two-minute change rather than a ticket. The chat interface ships in 12 languages and the agent follows the visitor's language, so the disclosure travels with them. Every conversation lands in the workspace inbox, which is the boring part of compliance: being able to show what your agent said, to whom, and when.
This is not legal advice and your counsel should have the final word on wording. But of all the AI regulation arriving over the next two years, this is the cheapest to satisfy and the most embarrassing to be caught missing. The deadline was three weeks ago.