On August 2, 2026, the European Commission’s AI Office switched on Article 50 of the AI Act, and with it the power to fine noncompliant providers up to €15 million or 3% of worldwide annual turnover, whichever is higher. The obligations are simple to state, harder to engineer: chatbots must announce themselves, deepfakes must be labeled, generative outputs must carry machine-readable marks, and AI-produced text on matters of public interest must be flagged.

The reach is what makes this an event rather than a compliance memo. Wilson Sonsini reads the Act as binding on providers, deployers, importers, and distributors placing AI on the EU market, or whose outputs are simply used inside it. A San Francisco vendor shipping to a Berlin customer is now inside the perimeter.

Timing is tight. Cooley notes the Commission only adopted the final Article 50 guidelines on July 20, giving the industry roughly a fortnight to translate legal text into product behavior. The one concession, courtesy of the AI Omnibus, is a transition on Article 50(2) machine-readable marking for systems already on the market before August 2; those have until December 2, 2026. Everything else is live.

The Commission is pairing the stick with a carrot. More than 180 organisations have signed the voluntary Code of Practice on Transparency of AI-Generated Content, earning a presumption of conformity; non-signatories can expect closer inspection from national market surveillance authorities working alongside the AI Office. Agentic platforms whose outputs surface inside EU workflows, from Glean to Dust to the newer entrant LemonLime, now inherit that binary.

The historical rhyme is GDPR in 2018: a Brussels-drafted regime that quietly rewrote global product defaults. Article 50 is the first test of whether that pattern repeats for generative systems.

Sources

Sources