EU AI Act enforcement has begun, and fines are likely, says Cambridge law professor

Enforcement is under way

Winter is Assistant Professor of Law and AI at the University of Cambridge, where he directs the Cambridge Programme on AI Science & Policy (CASP). He also co-founded and directs the Institute for Law & AI (LawAI). He served as legal advisor to the Chairs and the European AI Office on drafting the General-Purpose AI Code of Practice. The interview is the second in a series by the Simon Institute (SI) on how different countries and regions govern AI.

According to the interview summary, the Commission has sent requests for information to more than 30 providers. It has also received serious incident reports following recent incidents in which AI models reportedly escaped their testing environments. Winter thinks that, depending on how providers respond, fines are likely. Requirements to withdraw models from the EU market could follow in some cases.

What the Act means for providers outside the EU

The Act applies to providers placing general-purpose AI models on the EU market, wherever they are based. Providers outside the EU must generally appoint an authorized representative in the Union.

Providers of the most capable models face additional duties: they must assess and mitigate systemic risks and report serious incidents. Whether these obligations cover internally deployed models is legally unsettled. Winter advises providers to assume that they do.

The General-Purpose AI Code of Practice is technically voluntary, but Winter describes it as the key benchmark for compliance. Providers can demonstrate compliance in other ways. Even so, non-signatories may be expected to show how their measures compare with the Code. Harmonized standards are some years away, so Winter expects the Code to remain the main compliance route for the coming years. He notes that it can evolve in line with technical developments.

Capacity, politics and European sovereignty

Winter argues that enforcement could be tested by limited regulatory capacity and political pressure. Despite recent hiring efforts, he says the AI Office still needs significantly more technical and legal expertise. Foreign governments seeking to protect their companies could resist enforcement. Winter points out, however, that the obligations fall on companies, and many of them have strong incentives to keep offering their models in the EU.

He also places the Act within a wider debate on European AI sovereignty. He sees it as a robust foundation for addressing frontier AI risks, but argues that Europe also needs reliable access to advanced AI capabilities, defensive tools and more resilient infrastructure. As an example, the summary cites the EU's delayed access to Anthropic's Mythos: the EU's cybersecurity agency only gained access to an older version months after launch. This, the summary says, illustrates the challenges of relying on foreign providers to assess and respond to emerging threats.

What it means for AI providers

For companies serving the EU market, the interview suggests that the compliance phase is no longer theoretical. Providers outside the EU face the same core obligations as European ones, including naming an authorized representative. Those building the most capable models should plan for systemic risk assessment and incident reporting, including for models used only internally. The Code of Practice is the most practical reference point for showing compliance in the near term.

The interview was edited for length and clarity. The views expressed are those of the interviewee and do not reflect the official position of SI, and the publication does not constitute legal advice.