European Commission publishes guidelines for providers of general-purpose AI models

The European Commission published guidelines clarifying which providers of general-purpose AI (GPAI) models must comply with the EU AI Act's GPAI obligations and what that compliance involves. The guidelines arrived ahead of the AI Act's GPAI obligations entering into application on 2 August 2025.

What it is

The guidelines interpret the AI Act's requirements for GPAI providers, with a particular focus on Articles 53 and 55 covering transparency, copyright compliance and systemic risk mitigation. They are designed to complement the General-Purpose AI Code of Practice, a voluntary tool that explains how providers can fulfil their obligations. The Commission states the guidelines, though not legally binding, reflect its interpretation and will guide enforcement.

What it requires

A model is treated as GPAI if it is trained using more than 10²³ FLOPs and is capable of generating language, text-to-image, or text-to-video outputs. Any model trained using at least 10²⁵ FLOPs is presumed to have high-impact capabilities and may be classified as carrying systemic risk. Providers must notify the Commission within two weeks of reasonably foreseeing or reaching the systemic-risk compute threshold. A downstream actor that modifies a model becomes the new GPAI provider only if the modification's training compute exceeds one-third of that used for the original model. The guidelines also clarify under what conditions providers of open-source AI models are exempt from certain obligations. The Commission's enforcement powers, including fines, enter into application from 2 August 2026, and providers of models placed on the market before 2 August 2025 must comply by 2 August 2027. Noncompliance can result in fines of up to €15 million or 3% of global revenue, whichever is higher.

What security leaders should do

Security leaders should map which AI models their organization builds, fine-tunes or deploys against the FLOPs thresholds described above to determine whether GPAI provider obligations apply. Teams that fine-tune third-party models should track the compute they add relative to the original training run, since crossing the one-third threshold can shift provider status onto them. Organizations that rely on GPAI models with systemic risk should confirm their vendors have documented safety, security and incident-reporting processes, since those obligations extend through the model lifecycle. Compliance, legal and security functions should coordinate now given the enforcement timeline running through 2 August 2026 and 2 August 2027.