AUGUST 5, 2026 / ARTICLE 50

The deepfake label has arrived. Trust hasn’t.

Europe now requires disclosure for deepfakes and some AI content. The difficult question is no longer whether to label—but what the label can actually prove.

By Yifu HeAugust 5, 20267 MIN READ + 6 CASES

On August 2, the EU AI Act’s Article 50 transparency rules began to apply. Providers must make synthetic outputs detectable. Professional deployers must visibly disclose deepfakes and certain public-interest text. Those are different jobs—and collapsing them into one generic “AI-made” badge misses the point.

TWO LAYERS OF TRUTH

A mark for machines is not a warning for people

Article 50 separates infrastructure from interface. At the infrastructure layer, generative systems should attach interoperable, machine-readable signals to synthetic text, images, audio and video. At the interface layer, people encountering a deepfake should receive a clear disclosure no later than first exposure.

One helps platforms and tools trace provenance. The other helps a person interpret what is in front of them. A product can technically satisfy one while making the other almost invisible.

INTERACTIVE / LABEL OR NOT?

THE DISCLOSURE DESK

Six borderline cases based on the Commission’s guidance. This is an explainer, not legal advice.

CASE 01 / 06

POLITICS / VIDEO

A candidate appears to endorse a policy they never supported.

Their face and voice were generated to resemble a real press conference.

MAKE THE CALL

THE DEFINITION

Not every AI edit is a deepfake

The Act focuses on generated or manipulated image, audio or video that resembles existing people, objects, places, entities or events and would falsely appear authentic. Ordinary editing that does not substantially alter the input or its meaning is treated differently.

Satire and fiction are not a free pass. They receive a lighter disclosure regime designed not to interrupt the work. Public-interest AI text has another exception: visible disclosure is not required after human review or editorial control when someone accepts editorial responsibility.

THE PRODUCT DECISION

Compliance can still be designed to disappear

Google now places AI-ad disclosures inside a “How this ad was made” panel reached through a menu; local rules may make a label appear directly on the ad. Meta has signed the EU code while warning that too many labels could overwhelm people. Both positions reveal the real contest: not whether disclosure exists, but how much attention it receives.

A tiny badge, a buried panel and a full-screen warning can all contain the same words while producing completely different behaviour. Transparency is a product surface, not merely a line in a compliance checklist.

THE LIMIT

A disclosure explains origin. It does not establish truth.

A labelled clip may still be accurate. An unlabelled clip may still be fake. A criminal voice clone will not volunteer a warning. Labels help honest systems describe their outputs; authentication, provenance and editorial accountability are what make deception harder.

The label is not proof. It is a promise about process—and promises need an audit trail.

READ / VERIFY / UPDATE

PRIMARY SOURCES

01European Commission Article 50 guidelines ↗02Official transparency rules quick facts ↗03EU icons, examples and exceptions ↗04Meta signs the transparency code ↗05Google expands AI labels in ads ↗