Your AI disclosure is a design problem

Since 2 August 2026, if your product talks to people in the EU, you have to tell them when the thing talking back is a machine. That is Article 50 of the EU AI Act, and it reaches companies outside the EU whose systems are used inside it, which covers most British software with a European customer list.

Most teams will treat this as a legal task. Someone in legal writes a sentence, someone in engineering finds a place to put it, and the place is the footer. That is compliant on paper and useless in practice, because nobody reads a footer at the moment they are deciding whether to believe an answer. Disclosure is not a sentence you own. It is a decision about when the user finds out, and that is an interface decision.

Do we have to tell users they are talking to an AI?

If they are in the EU, yes, with one exception that is more interesting than the rule. Providers have to design systems so that people are informed they are interacting with an AI system, unless that is already obvious to a reasonably well informed person. Deployers pick up separate duties for deepfakes, emotion recognition, and published text on matters of public interest that no person has reviewed. The European Commission has issued guidelines and a code of practice covering all of it. Getting it wrong is expensive: fines run to 15 million euro or 3% of worldwide annual turnover, whichever is higher.

Now read the exception again. The law's own get-out is that it can be obvious. Obvious is not a legal category, it is a property of your interface. Whether a person can tell that a machine produced something depends on what the feature is called, how the output arrives, what sits next to it, and what happens if they accept it. A regulator asking whether your disclosure was necessary is, in effect, reviewing your design.

When is it obvious that something is AI?

A chat window with a cartoon face and a name is obvious. A button marked "summarise" that quietly rewrites a customer's words is not. Neither is an inbox that drafts replies in your own voice, or a dashboard where one figure is measured and the figure beside it is predicted.

The products most exposed here are the ones that did the tasteful thing and made the AI feel native. The better the integration, the weaker your claim that anyone could tell. That is an uncomfortable trade, and it belongs in a design review rather than a compliance review. The question is not where the notice goes. It is this: at which moments in this product would a reasonable person not know a machine was involved, and what do we show them there.

A disclosure that arrives after someone has already trusted the answer is not a disclosure, it is an alibi.

Where the disclosure belongs

Beside the output, not around it. Label the generated paragraph, the suggested reply, the predicted number. People read at the point of decision, so the label has to be at the point of decision.

Before the irreversible step, not after it. If the system is about to send, post, file or pay, that is the boundary between acting alone and asking first, and it should be drawn on purpose rather than inherited from whatever the framework did by default.

In the product's own voice. A warning in small grey italics reads as a hedge written by somebody else. The same fact, set in the product's own type and tone, reads as confidence. Identical information, opposite message.

If you already build a trust layer, with sources, confidence, an audit trail and a reversible action, none of this is new work. It is the work you were doing anyway, now with a date attached to it.

What about the UK?

There is no equivalent British statute. The UK has no cross-cutting AI Act, and the government's position has been that AI should largely be regulated at the point of use by existing regulators such as the ICO and Ofcom. A product with no EU users sits outside Article 50 entirely.

That is a thin reason to do nothing. The regulation is the deadline, not the argument. The argument is that a product which marks its machine output plainly is easier to sell into a cautious sector than one that blurs it, and far easier to defend on the day the output is wrong. It is the same trade as publishing your pricing: the thing you were nervous to say turns out to be the thing that qualifies the right people in.

The compliance will happen either way. The only real choice is who owns it: the people who write policy, or the people who design the moments where a user decides what to believe. The first produces a document. The second produces a product that can be checked.