Telling People They’re Talking to a Machine Is Now a Design Problem

For most of the past three years, whether to disclose that a customer is talking to an AI has been treated as a matter of taste. Some teams put a small robot icon in the corner. Some open with a line about being a virtual assistant. Plenty do neither and rely on the interaction being short enough that nobody wonders.

That period is ending on a specific date. From 2 August 2026, the transparency obligations in Article 50 of the EU AI Act apply, and the question moves from brand preference to product requirement for providers of in-scope AI systems that interact directly with people and fall within the EU AI Act’s territorial scope.

The Obligation in One Sentence

Providers of AI systems designed to interact directly with people must build them so that those people are informed they are interacting with an AI, unless that fact is obvious to a reasonably well-informed, observant, and circumspect person.

Two things in that sentence do most of the work. The obligation sits with the provider, meaning it has to be designed into the system rather than bolted on by whoever deploys it. And the exception, that it might be obvious, is narrower than teams hope, because it is judged from the perspective of a reasonable person rather than from the perspective of the engineer who built the thing.

Who This Actually Catches

The scope is narrower than the panic and broader than the complacency.

Systems that qualify are those designed for a genuine two-way exchange with people, where the AI itself communicates directly rather than through a human intermediary, and where the counterpart is a natural person, whether a consumer, a professional or any other user. Support chatbot, voice agents, virtual assistants and avatars all sit squarely inside.

What falls outside is anything running in the background. Machine-to-machine communication, systems that merely collect data or return automated responses without genuine exchange, and anything with no direct contact with a person are not caught by this particular obligation.

The Extraterritorial Part Americans Keep Missing

Here is the detail that catches US companies who assume this is a European problem for European vendors.

Providers established or located outside the EU are subject to the AI Act where the output of their system is used in the EU. A US company shipping conversational ai into a product with European users is in scope regardless of where the servers or the cap table sit, and the provider must meet the applicable transparency requirements before placing the system on the EU market or putting it into service, where those provisions apply.

If your product falls within the AI Act’s territorial scope, the obligation can apply to you as the provider; the presence of an EU distributor does not automatically shift that responsibility away from you.

Disclosure Has a Deadline Within the Conversation

The requirement is not merely that disclosure exists somewhere. It has to arrive at the right moment and in the right form.

The information must be given clearly and distinguishably, at the latest at the time of the first interaction or exposure, and it must meet applicable accessibility requirements. That rules out a mention buried in terms of service, a disclosure that appears only after the user has asked a question, and anything that fails for a screen reader user.

For a voice product, that means disclosure should occur at the start of the first interaction, before the substantive exchange begins.

What the Commission Has Actually Published

None of this is being inferred from the statute alone, which is worth knowing because guidance materially changes what compliance looks like.

The European Commission’s own materials on transparency obligations under Article 50 set out the cumulative criteria described above and confirm the extraterritorial reach for non-EU providers. The Commission has also published Guidelines on transparency obligations covering interactive and generative systems and deepfakes, alongside a Code of Practice on Transparency of AI-generated Content. For the separate marking and labelling obligations covering AI-generated or manipulated content, adherence to the Code of Practice is one way of demonstrating compliance; organisations that do not follow it must demonstrate compliance through other adequate means. For direct AI interaction under Article 50(1), providers determine appropriate measures while taking the Commission’s guidelines into account. For the interaction disclosure specifically, providers and deployers are left to determine adequate measures themselves while taking the guidelines into account.

Enforcement runs through national market surveillance authorities and the AI Office, with the European Data Protection Supervisor covering EU institutions acting as providers or deployers.

This is general information rather than legal advice, and obligations turn on specifics. Anyone deploying into the EU should get qualified counsel rather than relying on a summary.

Good Disclosure Is Not a Disclaimer

Now the part that regulation does not solve for you.

There is an enormous difference between a compliant disclosure and a good one. “This assistant is powered by AI” satisfies the requirement and tells the user almost nothing useful. What users actually want to know at the start of an interaction is what the system can do, what it cannot, and how to reach a person.

A useful disclosure can combine three things: an unambiguous statement of what the user is talking to, a one-line description of its scope, and an immediately available route to a human. That is more useful than any icon and it happens to be cheaper than the support tickets generated by users who discovered the truth halfway through.

The Trust Argument Is Stronger Than the Compliance One

Executives frequently assume disclosure suppresses engagement. There is also a product-design argument for disclosure. Users who know they are interacting with AI can calibrate their expectations, ask questions differently and escalate when the system is unable to help.

Users who know they are talking to a system adjust how they speak. They ask more direct questions, they skip the social preamble, and they escalate earlier when the system is out of its depth. All of that improves resolution rates. Users who believe they are talking to a person and then work it out do the opposite: they feel deceived, they lose confidence in everything the system already told them, and the interaction usually ends with an angrier human conversation than would otherwise have happened.

Concealment buys you a slightly smoother first minute at the cost of the whole relationship when it fails.

What to Do Between Now and Your Next Release

Four things, none of which require a legal department to start.

Inventory every surface where your systems talk directly to people, including the ones nobody thinks of as products, such as internal tools that reach contractors. Establish for each whether a reasonable user would already find the AI involvement obvious, and assume not. Move disclosure to the first interaction and check it works for assistive technology. And write down which measures you chose and why, because demonstrating compliance is a documentation exercise as much as a design one.

The Era of Ambiguity Is Closing

There was a window in which not saying was a defensible product choice, and it is shutting for reasons that go beyond one regulation. Users are getting better at spotting synthetic interaction, competitors are disclosing, and the reputational cost of being perceived as concealing AI involvement can be significant, particularly as users become more familiar with synthetic interactions.

Building the disclosure in properly now is considerably easier than retrofitting it under a deadline, and the teams who treat it as an interface problem rather than a legal one tend to end up with a better product either way.