Europe Forced Authentic-Looking Deepfakes to Wear Labels

The carve-outs matter. Fantastical or obviously unreal content sits outside. AI text that a human with real editorial responsibility has reviewed can skip the public-interest label. Artistic and satirical work still needs a disclosure, but one that does not wreck the experience. Systems already on the EU market before 2 August get until 2 December 2026 for the machine-readable marking piece. Deepfakes made before the deadline need no retroactive sticker, though the Commission encourages one.

Penalties match the ambition: up to €15 million or 3% of worldwide annual turnover, whichever is higher. National market surveillance authorities lead most of the chase. The EU AI Office watches systems under its own remit. The Commission has published guidelines and a Code of Practice on marking, with optional icons for deployers who want a common visual language.

The philosophical wager is larger than compliance checklists. After Flávio Bolsonaro's synthetic father walked into a Brazilian convention fight, the question stopped being whether politics would invent faces. It became whether voters would know they were looking at invention. Europe's answer is ambient confession: make the synthetic declare itself before trust finishes rotting.

Enforcement will decide if the confession is real. Watermarks strip. Metadata evaporates. Machine-written prose resists reliable detection. Global platforms will likely ship one labeling stack worldwide rather than maintain a Europe-only fork, so Brussels may export its epistemology the way it once exported privacy rules. That only helps if detectors keep pace with generators.

For now, the law is a bet that shared reality needs paperwork. Europe chose the stamp over the ban for everyday synthetic media, and left the harder bans on non-consensual intimate deepfakes for a later date. Whether a label can hold a public's attention once every feed wears one is the next fight.