On 2 August 2026 the EU AI Office acquires enforcement powers over providers of general-purpose AI models. The powers are worth reading article by article, because they are considerably sharper than the phrase “enforcement powers” suggests.
The powers
Article 91 lets the AI Office demand technical documentation, training-data summaries and model reports. Article 92 lets it commission independent evaluations, including with API and other technical access to the model. Article 93 lets it require mitigation measures, restrict a model, or require its withdrawal or recall from the market. Article 101 sets fines at up to 3% of global annual turnover.
Article 92 is the one to dwell on. A right to commission an independent evaluation with technical access is a very different instrument from a right to request a document. It is the difference between reading what a provider says about a model and testing the model.
The transparency side
Article 50 applies to four situations: systems interacting directly with people; the generation of synthetic content; emotion recognition and biometric categorisation; and deepfakes or AI-generated content concerning matters of public interest. Most legal-sector deployments touch at least the first two.
Still in motion
Several pieces were still open at the start of June. The Commission ran a targeted consultation on draft guidelines for classifying high-risk AI systems, closing 23 June 2026. The AI Office published a FAQ on general-purpose AI models, compiled from questions asked during AI Pact webinars – a more practical document than its status suggests. Regulatory sandboxes moved to August 2027. At national level, the Polish parliament’s Committee on Digitalization, Innovation and Modern Technologies adopted a draft AI bill, proposing a domestic oversight body, the Commission for the Development and Security of Artificial Intelligence (KRiBSI).
The national layer deserves attention from anyone outside the Union. The AI Act sets the obligations, but the body a provider actually deals with is usually a national one, and those bodies are being designed right now.
Drawn from The EU AI Act Newsletter #103, 1 June 2026.