Since August 2, 2026, the European Commission has received official authority to impose penalties on suppliers of general-purpose artificial intelligence (GPAI) models for non-compliance with the provisions of the Artificial Intelligence Act (AI Act). This stage marks the transition from declaring requirements to their actual enforcement, as the basic obligations for GPAI developers entered into force a year earlier. Companies are required to provide detailed technical documentation, disclose the necessary information to downstream providers, and strictly comply with European Union copyright laws. Suppliers of the most powerful systems that pose systemic risks are additionally required to notify the European Office for Artificial Intelligence and implement mechanisms for assessing and minimizing potential threats.

Financial penalties for violations of the regulations are differentiated depending on the severity of the offense. For suppliers of GPAI models who avoid providing documentation or refuse to open access to the model for evaluation, fines of up to 15.0 million euros or 3% of global annual turnover are provided. At the same time, strict transparency requirements have come into force: interactive systems and chatbots are required to explicitly notify users of the fact of interaction with artificial intelligence, and any generated or modified content, including deepfakes, must be marked with visible and machine-readable markings. For systems released to the market before August 2, 2026, a transition period is provided for the implementation of labeling mechanisms until the end of December this year.

To help companies adapt to the new standards, a voluntary Code of Practice has been developed covering transparency, copyright, and data security and protection issues. Leading technology corporations, including Anthropic, Google, IBM, Microsoft, Mistral AI and OpenAI, have already signed it, while xAI has agreed only to the security provisions, which will require it to independently prove compliance with the remaining criteria. The application of full-fledged rules to high-risk AI systems and algorithms embedded in physical products such as medical devices or elevators has been postponed until 2027-2028. This postponement, related to the entry into force of the Digital Omnibus Act, is intended to provide national regulators with additional time to develop standards and configure competent authorities.