EU’s AI Act Outreach Tests Whether Labs Can Prove Their Safety Claims

Article 91 sets out the Commission’s power to request documentation and information from GPAI-model providers. Source: EUR-Lex.

A request is not yet a formal investigation

The distinction between outreach and a formal statutory demand is central. Under Article 91 of the AI Act, the Commission may request GPAI documentation prepared under Articles 53 and 55, or additional information necessary to assess compliance. The European AI Office, the Commission body that carries out these Chapter V GPAI supervisory tasks, may initiate a structured dialogue before sending that request; a formal request must identify its legal basis and purpose, state what is required, set a deadline and flag the applicable fines.

The Commission has not published the letters or said whether the reported communications meet that standard. A legal analysis of the reported action therefore cautions that it cannot establish an Article 91 request, a formal investigation or an infringement finding.

If information obtained through Article 91 is insufficient, the AI Office may, after consulting the European AI Board, evaluate a model. The Commission can seek access through APIs or other appropriate technical means, including source code. In specified cases of intentional or negligent non-compliance—including failure to comply with an Article 91 request, a remedial measure or evaluation access—the maximum penalty is €15 million or 3% of the prior year’s worldwide annual turnover, whichever is higher. Those powers describe a possible escalation path; they do not show that it has begun for any recipient of the reported correspondence.

Article 55 lists evaluation, risk-mitigation, incident-reporting and cybersecurity duties for systemic-risk GPAI-model providers. Source: EUR-Lex.

The compliance question is whether a provider can reconstruct its decisions

All GPAI providers have baseline duties including technical documentation, information for downstream providers, a copyright-compliance policy and a sufficiently detailed public summary of training content. Some documentation and downstream-information duties can be exempted for qualifying free and open-source models, but that exception does not remove copyright or training-summary duties and does not apply to GPAI models with systemic risk.

The more demanding category is GPAI with systemic risk, not the AI Act’s separate category of high-risk AI systems used in particular contexts such as employment or education. A GPAI model is presumed to have high-impact capabilities when its cumulative training compute exceeds 10²⁵ floating-point operations, though the Commission may also classify a model on equivalent capability or impact grounds. The threshold is a presumption, not evidence that any named model or company is part of this outreach.

For systemic-risk GPAI models, the Act requires state-of-the-art evaluations, documented adversarial testing, assessment and mitigation of Union-wide systemic risks, serious-incident monitoring and reporting, and appropriate cybersecurity for both the model and its physical infrastructure. That is why the reported questions about evaluator access and post-release monitoring matter: a public safety policy alone would not establish that a provider performed the tests, documented the results, considered outside advice and corrected failures.

The EU’s GPAI code of practice can offer a recognised route for showing compliance, but it is voluntary and does not create the obligations. Signing it, declining to sign it or previously discussing an incident with officials is not proof of receiving these requests.

What would resolve the central question

The next useful evidence would be a Commission disclosure of the communications’ legal basis and scope, a recipient’s confirmation, or a published escalation decision. Without it, treating a roster of high-profile labs as a list of targets would outrun the evidence.

The larger question is whether the AI Office can turn broad safety, security and copyright duties into focused requests, and whether providers can answer with a traceable record across a model’s lifecycle. The reported outreach starts that test. It does not yet reveal who passed, who failed or whether the Commission will need to use its stronger evaluation and remedy powers.

Judge lets Claude Max claims over ‘5x or 20x more usage’ proceed

A federal judge found Anthropic’s Claude Max usage advertising plausibly misleading because subscribers could read the multipliers as monthly promises. Contract and misrepresentation claims survive, but several other claims were dismissed with leave to amend.