Since August 2, the European Commission has officially gained the authority to impose financial penalties on providers of general-purpose AI (GPAI) models for violating the AI Act. This marks a shift from declarative norms to actual enforcement — a key stage in shaping the global regulatory landscape.

New transparency rules: what has changed for users

Simultaneously, transparency requirements come into force. Chatbots and interactive AI systems are now obliged to directly inform users that they are interacting with a machine. Deepfakes and any content created or modified by AI must receive machine-readable labels to facilitate their identification. This also applies to emotion recognition systems and biometric categorization — their use must also be disclosed.

Mechanics of fines: two liability regimes

The key innovation is a differentiated approach to penalties. For providers of GPAI models, a separate regime under Article 101 applies: the maximum fine reaches €15 million or 3% of worldwide annual turnover. Sanctions are applied for refusing to provide documentation, submitting inaccurate data, or failing to meet risk assessment requirements.

However, for prohibited practices — such as social scoring or manipulative systems — a stricter regime under Article 99 is provided. Here, the ceiling rises to €35 million or 7% of global turnover. For providing incomplete information to national regulators — a fine of up to €7.5 million or 1% of turnover. For small and medium-sized businesses, the lower of the two thresholds applies, demonstrating the regulator's attempt to maintain a balance between innovation and control.

Transition periods and the voluntary code

For companies that entered the market after August 2, 2025, the obligations are already in effect. Other providers are granted a deferral until August 2, 2027. A separate transition period until December is established for labeling content created by generative systems placed on the market before August 2, 2026.

Notably, giants such as Google, Microsoft, OpenAI, and Anthropic have signed the voluntary Code of Practice. However, xAI agreed only to the chapter on safety, effectively leaving it outside the frameworks of transparency and copyright — this creates a precedent for a selective approach to regulation.

High-risk systems: pause until 2027

Rules for high-risk AI systems will only come into force on December 2, 2027, and for AI in physical products (medical devices, toys, elevators) — on August 2, 2028. The European Commission has also launched tools for filing complaints and a separate channel for whistleblowers connected to GPAI model providers.

My analysis: Europe is deliberately postponing the most burdensome norms for businesses, giving the industry time to adapt. However, introducing fines for opacity is a signal: the era of AI self-regulation is over. Now, compliance with the AI Act becomes a matter of financial sustainability, not just reputation. Companies, especially those with a global presence, will have to revise their compliance strategies in the coming quarters.