The AI Act stopped being a legal problem and became a UI requirement
By Allan Leone on
Article 50 of the EU AI Act became enforceable on 2 August. The obligations it creates are not paperwork. They are screens somebody has to design, and most teams have not drawn them yet.
The transparency obligations in Article 50 of the EU AI Act became generally applicable and enforceable on 2 August 2026. I have watched a lot of teams file this under legal review. It is not a legal review problem. Every one of these duties resolves to something a user has to see, which means it resolves to something a designer has to draw.
The scope is wider than most people assume. It applies whether or not your system is classified as high risk, and it applies to systems already on the market, not just new ones.
Four things you now have to show
- That the person is interacting with an AI system, when that is not already obvious.
- That content was generated or manipulated by AI, in a machine-readable way.
- That emotion recognition or biometric categorisation is being applied to them.
- That a deep fake, or AI-generated text published on a matter of public interest, is what it is.
Read that list as a design brief and it is four disclosure patterns, each with its own placement, timing and persistence question. None of them are hard. All of them are easy to get wrong in a way that either buries the disclosure or wrecks the flow.
The failure mode is the modal
The first instinct is always a dialog on first load. It is the worst option available. It interrupts before the user has context for what is being disclosed, it gets dismissed reflexively, and it discloses once for a session in which the AI might appear in six different places.
Disclosure works better attached to the thing it describes. A label on the generated image. A persistent marker on the assistant surface. A line in the message composer rather than a banner above the page. The user reads it at the moment it means something.
Machine-readable is the part that gets skipped
The obligation on generated content is not only that a human can tell. It has to be marked in a machine-readable format. That is a pipeline change, not a design change, and it lands on whoever owns the asset export. If your generated images leave the system without provenance metadata, the visible badge in the interface does not cover you.
What it costs to get ready
For the products we looked at, this was a week of work, not a quarter. An inventory of every surface where a model touches the user, a decision per surface about what has to be disclosed, and a component to do it consistently. The expensive version is the one where four teams each invent their own badge.
The penalties are up to 15 million euro or 3 percent of worldwide annual turnover, so the case for doing it properly writes itself. But the reason to do it well is that a product which is straight about what is a machine is easier to trust, and that was true before there was a regulation attached to it.
Tags: compliance, product, ai