Europe Officially Enforces the AI Act, Opening the Door to Global Fines

On August 2, 2026, the European AI Office activated its full enforcement powers under the AI Act, ending the grace period one year after the law's adoption and opening the door to retroactive fines for violations traceable back to August 2025.

Transparency obligations stipulated in Article 50 take effect immediately: chatbots and voice agents are required to disclose their artificial nature at the start of every user interaction, and regulations require marking AI-generated or modified content with machine-readable watermarks. Fines range between 15 million euros or 3% of annual global turnover for general-purpose model violations, rising to 35 million euros or 7% in the case of specifically prohibited practices.

The eight models whose computing capacity exceeds the threshold of 10²⁵ floating-point operations have been subject to monthly systemic risk evaluations since August 2025. While the deadline for conformity assessment of high-risk AI has been extended to December 2, 2027, transparency obligations remain active now without delay. The adaptation period for marking deepfake content ends on December 2, 2026.

Legal and technology experts believe this enforcement radically changes the equation: every company in the world—regardless of its geographic location—that places its models on European markets or trains them on European data is legally exposed today. The European AI Act has transformed from a legislative document into an actual sanctioning tool—and this moment is the first time it carries real weight.

What do these terms mean?

EU AI Act: A European law that classifies artificial intelligence systems according to their risk level and obligates companies to meet requirements proportionate to the level of risk—applying to any entity offering its services in EU countries regardless of its headquarters location.

General-Purpose AI (GPAI) Models: Large-scale AI models—such as ChatGPT, Claude, and Gemini—designed to perform multiple and diverse tasks, subject to special requirements when their computing capacity exceeds thresholds set by regulations.

Systemic Risk: The risks that a large-scale AI model may pose to society as a whole, such as disrupting critical infrastructure or spreading widespread disinformation that is difficult to halt.