The European Commission has published an account of how independent outside scrutiny of powerful general-purpose AI models might work. The report comes from a workshop held on July 15, with 20 organisations selected and 19 taking part. Its publication on Wednesday makes the discussion newly visible; it does not mean the EU has adopted a scoring system for evaluators or completed new safety tests on any model.
The European AI Office report describes possible requirements for organisations that could evaluate general-purpose AI models. Participants discussed qualifications, relevant skills and domain expertise, along with the independence of evaluators from the providers they assess. The report also records possible safeguards against managerial, financial or operational influence.
Those questions matter because a nominally external review can still depend on the company being reviewed. A provider might pay for an assessment, supply technical access and define parts of the testing environment. Independence therefore has several dimensions: who selects and pays the evaluator, whether it can choose methods, what evidence it can inspect, and whether it can publish or report uncomfortable findings. The workshop report identifies some of these tensions, but it does not resolve them into a final standard.
Compliance is taking shape
The discussion sits alongside the EU’s broader framework for general-purpose AI. The Commission says its Code of Practice is a voluntary tool to help providers demonstrate compliance with the rules, with a separate Safety and Security chapter for providers of models presenting systemic risks. “Voluntary” describes adherence to the code; it does not make the underlying legal obligations optional for providers covered by the law.
The Commission’s guidance for providers sets out how it interprets obligations under the AI Act. The workshop report concerns a narrower institutional question: what makes an outside evaluator qualified and sufficiently independent when external evaluation is required. These materials are related, but a discussion of evaluator criteria should not be mistaken for a determination that a particular model has systemic risk or has passed an assessment.
A credible evaluator regime could give regulators and the public more confidence that safety claims have been examined beyond a provider’s own teams. It could also create a market in which a handful of firms control scarce expertise, or in which limited access to model weights and training information constrains what reviewers can verify. The July participants’ views, as summarised by the Office, offer an early map of the questions. They are not evidence that those practical problems have been solved.
The next test is whether the Commission turns the workshop’s ideas into clear, enforceable criteria and gives evaluators enough access to do meaningful work. Until then, Europe has a set of design proposals and existing legal duties, not a functioning independent certification system.
