Skip to content
NewsRegulation

EU AI Act enforcement went live August 2, giving the AI Office direct power over GPAI providers

· by Pondero Newsdesk

The short version

The European Commission's AI Office began enforcing the EU AI Act on August 2, 2026, activating rules for general-purpose AI model providers, prohibited AI practices, and chatbot transparency.

EU AI Act enforcement went live August 2, giving the AI Office direct power over GPAI providers

The European Commission's AI Office gained active enforcement authority over general-purpose AI model providers on August 2, 2026. The same date activated transparency rules for chatbots and deepfakes. For OpenAI, Anthropic, Google, and every other signatory to the GPAI Code of Practice, the compliance window is closed.

What changed on August 2

Per the Commission's own press release, the AI Office and 24 national competent authorities started enforcing three clusters of rules simultaneously.

The first covers GPAI model obligations. All providers must make technical documentation available to downstream system builders and operate a copyright compliance policy. Providers of the most advanced models carry additional burdens: Article 51 of the AI Act designates any model trained on more than 10^25 floating point operations as a "GPAI model with systemic risk," and those providers must conduct ongoing safety evaluations and give the AI Office access to their models on request.

The second cluster covers prohibited AI practices: social scoring, subliminal manipulation, real-time remote biometric surveillance, and predictive policing based solely on profiling.

The third covers chatbot and synthetic-content transparency. Any interactive AI system must tell users it is not human. Deepfakes must carry machine-readable marks. Over 180 organizations signed the voluntary Code of Practice on AI-generated content transparency ahead of the date.

Who it covers

The GPAI Code of Practice lists the current signatories: Amazon, Anthropic, Google, IBM, Microsoft, Mistral AI, and OpenAI signed all three chapters. xAI signed only the Safety and Security chapter, meaning it must demonstrate compliance with the transparency and copyright chapters through other means.

Signing the code is voluntary but gives providers a defined path to compliance. Non-signatories must prove they meet the same obligations through other methods, and the AI Office can investigate either group. Enforcement powers for the Office include sending formal requests for documentation, conducting model evaluations through independent experts, restricting public model availability if warranted, and imposing fines. Penalties for GPAI model violations can reach 15 million euros or 3 percent of worldwide annual turnover, whichever is larger. Violations of the prohibited practices rules carry a ceiling of 35 million euros or 7 percent of turnover.

China's companion AI rules also in early enforcement

China activated its own companion AI framework on July 15. Per IAPP, the Cyberspace Administration of China's Interim Measures for the Administration of Anthropomorphic AI Interaction Services require chatbot companion providers to detect emotional distress, intervene in addictive use patterns, and bar minors from persuasive AI avatars. IAPP reports the CAC issued fines to 12 companies in the first weeks of enforcement, though the CAC has not published a public enforcement ledger.

Both regulatory actions center on disclosure and user protection rather than capability restrictions. Neither China nor the EU has announced a ban on any existing frontier model.

Why it matters

The AI Office's investigative powers activate without requiring a complaint: it can initiate evaluations on its own and appoint independent technical experts to assess any GPAI model it classifies as high-risk. Teams building on covered GPAI models in the EU now carry deployer obligations tied directly to how the underlying model was documented. The 10^25 FLOP threshold is the key operational gate; any model crossing it is presumed to carry systemic risk under Article 55 regardless of the provider's own self-assessment.

What to watch next

The AI Office has not yet announced a first formal documentation request targeting a named frontier model provider. Its next visible action will define whether August 2 marks a soft opening or an immediate audit posture. IAPP's analysis notes the companion-AI rules could be extended to general generative model providers in future CAC amendments, though no draft timeline has been published.

Sources