EU AI Office Gains Full Penalty Powers Over General-Purpose AI; Article 50 Chatbot Disclosure Becomes Enforceable

On August 2, 2026 the European Commission’s AI Office activated its full enforcement toolkit over general-purpose AI model providers, converting obligations that had technically been in force since August 2025 into penalties that can actually be imposed. The office can now demand documentation and information under Article 91, conduct its own model evaluations under Article 92, compel corrective measures under Article 93, and fine under Article 101 — up to the greater of €15 million or 3 percent of worldwide annual turnover for GPAI violations, and up to €35 million or 7 percent for prohibited AI practices. The providers in scope are overwhelmingly American: OpenAI, Google, Meta, Anthropic. The same date made Article 50 transparency obligations enforceable — an AI chatbot, voice agent, or interactive system deployed in the EU must tell users at the start of an interaction that they are dealing with a machine, synthetic content must carry machine-readable marking, and deepfakes must be labeled. Models first placed on the market before August 2, 2025 have until August 2, 2027 to reach full compliance.

The shape of what activated and what did not is the story. The obligations that took effect on schedule are the disclosure-and-transparency layer: tell people it is an AI, mark the synthetic content. The obligations covering AI systems that make consequential decisions about people — employment screening, credit scoring, law enforcement, critical infrastructure — were pushed to December 2027 by the AI Act Omnibus, Regulation (EU) 2026/1744, which entered force barely a week before the enforcement date (2026-07-27–eu-ai-omnibus-delays-high-risk-ai-deadline-to-december-2027). The facial-recognition prohibitions held to their original timeline (2026-07-25–eu-ai-act-full-enforcement-facial-recognition-bans). What arrived on time is the layer that regulates whether users are told; what slipped sixteen months is the layer that regulates whether a machine may decide.

Against this, the American direction is the reverse on both axes. There is no federal equivalent of the AI Office and no penalty authority resembling Article 101; the live federal legislative effort points toward a three-year preemption of state AI regulation (2026-06-04–great-american-ai-act-discussion-draft-three-year-state-preemption), and the surveillance model the EU prohibits is being cleared for streamlined federal procurement on this side (2026-07-08–clearview-fedramp-cbp-facial-recognition-all-travelers). The divergence now has an enforcement asymmetry attached: the same firms face a regulator with subpoena, evaluation, and fining power in one market and, in their home market, an active effort to bar the states from regulating them at all.