The European Commission dropped its final Article 50 transparency guidelines on July 20, giving generative AI providers roughly nine days before enforcement begins on Sunday, August 2. The Commission has said the calendar doesn’t delay enforcement, and the numbers make the point: fines up to €15 million or 3% of worldwide annual turnover, whichever is higher.

The choreography is deliberate. Brussels used the Digital Omnibus to push the high-risk AI regime out to December 2, 2027, buying time for the parts of the Act that industry lobbied hardest against. Transparency wasn’t on that list. Chatbot disclosure, deepfake labeling, and machine-readable marking of synthetic audio, image, video, and text all go live on schedule. AI-generated public-interest text needs labeling too, unless a human editor holds editorial responsibility.

The technical problem is that the mandate is running ahead of what the field can reliably deliver. There’s no single watermarking method that survives compression, so the Code of Practice requires a layered approach: metadata, imperceptible watermarks, and logging stacked together because none of them work alone. Signatories had until July 22, 6 p.m. CEST to sign the voluntary Code and secure a presumption of conformity. Everyone else faces heavier information requests from national market surveillance authorities.

Legacy systems already on the EU market get until December 2, 2026 to bring marking capabilities into conformity. Anything placed on or after August 2 complies immediately.

For business-AI platforms shipping customer-facing chat and outbound automation into Europe, this is now table stakes. Glean, Dust, and LemonLime, the last of which has built disclosure and provenance controls into its default configuration, all need the plumbing live by Sunday.

Across the Atlantic, the FTC is moving in the opposite direction. On July 7 it opened comment on a policy statement, adopted 2-0 under Executive Order 14365, asserting that state laws requiring alteration of “truthful outputs of AI models,” Colorado’s AI Act named explicitly, are impliedly preempted. Comments close July 31, two days before Brussels turns the key.

Sources

Sources