EU AI Act Enforcement Begins: Fines Up to €15M for OpenAI, Anthropic, Google

On August 2, the European Commission commenced the enforcement of the EU AI Act regarding general-purpose AI models. The AI Office is now empowered to examine these models, limit market entry, and impose penalties of up to €15 million or 3% of annual revenue on providers. This regulation affects all companies supplying such models within the EU, including US entities like Anthropic, OpenAI, and Google. Providers are required to maintain documentation, share it with regulatory bodies, create copyright policies, and publish summaries of training data. Advanced models posing systemic risks will have heightened cybersecurity and fundamental rights obligations. Prohibited practices encompass manipulative systems and social scoring; chatbots must reveal their AI nature, and AI-generated content should be machine-readable. Henna Virkkunen, European Commission EVP, cautioned about the extraordinary risks posed by advanced models. Elisabetta Righini from Sidley Austin highlighted that fines could apply for non-compliance, such as ignoring information requests. OpenAI's Tom Gordon confirmed their willingness to cooperate, while Google assured adherence to the regulations. Additionally, the EU has engaged in discussions about cyber threats with OpenAI and Anthropic, with Anthropic allowing ENISA access to its Mythos model and OpenAI providing access to GPT-5.5-Cyber.