Design AI disclosures that users notice and regulators accept

The failing disclosure arrives after the user has already trusted the interface. For EU-facing AI products, the timing is fixed: the EU AI Act's Article 50 transparency obligations begin applying on 2 August 2026.

Under Article 50, providers must ensure that AI systems interacting directly with people make clear they are AI, unless a careful, knowledgeable user would already recognise that.

Put the label where the user first meets the system

A first-interaction disclosure is a design decision. If a user can act before seeing the AI marker, it has missed the moment that matters. The marker must meet the user before trust forms. The redesign is simple: show the marker before the user can act, keep it visible while the session is active, and repeat it where generated output appears.

The obvious-exception test is where many teams get overconfident. A system may carry a familiar name, but users still need a clear signal that the current surface is machine-generated. If you rely on the exception, write the reason down and test it with a user who has not been told the answer.

Make the checklist the release gate

Run this against every AI surface before launch: image, audio, video, and any generated text that leaves the product.

• First-interaction disclosure: can a new user act before seeing the AI marker?
• Persistent label: is the AI marker still visible while the session is active?
• Pre-interaction notice: does the user get a warning before generated output is created, sent, or downloaded?
• Accessible and distinguishable placement: can a user meet the label without extra steps?
• Obvious-exception test: is there a written reason for relying on the obvious exception?
• Machine-readable marking: can the generated output be verified as machine-readable and detectable?
• EU icon selection: does the chosen icon match the actual degree of AI involvement?

Each item is a quick test because it forces a visible or recorded signal: a label, a notice, a machine-readable signal, or a documented exception. If the answer is that it will be added later, the surface is not ready.

The early items protect the entry moment; the middle items protect users who cannot rely on default cues; the final items protect the record after the user leaves.

Generated content needs visible and machine-readable layers

Users need to see that output is artificial. Machines need to detect it later. Under Article 50, deployers must disclose that image, audio, or video output is artificial when the system generates or manipulates content that amounts to a deepfake.

Generative AI systems already placed on the market before 2 August 2026 have a grace period until 2 December 2026 to make their output marking machine-readable and detectable.

The visible layer should sit near the output, not buried in a separate view. The Commission has published three free EU icons for marking AI-generated material, covering simple disclosure, wholly AI-made output, and partly AI-altered output, with no attribution required.

Pick the icon that matches the actual process, not the marketing claim. The icon set gives designers a way to make that signal legible without inventing a new visual language.

Text gets a narrower exception. For public-interest text, the duty does not apply when a human has reviewed or editorially controlled the text and a person or entity is responsible for its publication.

That exception should be recorded, not assumed. If a human review step can be shown, the label can be lighter. If it cannot, treat the text like any other generated output.

Do not let the label become a decoration

A disclosure that is too small, too faint, or too far from the interaction surface will pass a legal review and still fail the user. The label needs contrast, proximity, and a place in the interface that users can reach without extra steps. A label that is only visible to some users is not enough. If the label appears only after the user has already asked a question, the design has hidden the thing it was supposed to reveal.

Some surfaces add a further duty. Those who deploy systems for emotion detection or biometric categorisation must tell affected people and process the related personal data in line with GDPR.

Disclosure alone will not cover it. Put the notice where the exposed person first sees the result, and keep the data-processing question in the same design review. Notice and data-processing belong to the same team that ships the interface.