The Scientific Panel of Independent Experts: What Is It And How Does It Work?
AI Act Adam Radivojevic
The EU just picked the 60 experts who'll police powerful AI models under the AI Act. They can flag risky systems straight to Brussels, and that has real teeth.
Based on reporting by AI Act, Adam Radivojevic — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Brussels quietly assembled a watchdog with actual bite on June 1, 2026, when the European Commission published the roster of its new Scientific Panel of independent experts. Sixty names, drawn from at least 80% EU/EFTA/EEA countries, each guaranteeing at least one seat per member state and capping any single country at three. It's the kind of bureaucratic detail that sounds boring until you realize what these people are actually empowered to do.
The panel exists to backstop the AI Office as it tries to enforce rules on general-purpose AI models, the GPT-4s and Geminis of the world. Under Article 68 of the AI Act, these experts advise on how to classify models, help build evaluation benchmarks, and can even conduct their own model evaluations under Article 92. But the sharpest tool in the kit is the qualified alert under Article 90. If the panel suspects a model poses a concrete Union-level risk, or crosses the systemic-risk threshold, currently presumed at 10^25 floating-point operations of training compute, it can flag the Commission directly. That alert can trigger information demands, formal evaluations, and eventually a systemic-risk designation under Article 52, which saddles a provider with extra obligations around incident reporting and cybersecurity.
The process behind picking these 60 people took almost a year. The Commission opened its call for expressions of interest back in June 2025, closed applications in September, and spent the following months vetting candidates for independence, impartiality, and up-to-date technical chops. Every expert now has to file a public declaration of interests, and the AI Office is on the hook for actively managing conflicts, because letting industry insiders quietly grade their own homework would defeat the entire point.
What's notable is how much procedural friction is baked in. A qualified alert needs simple majority approval among panel members. A request for information to a provider needs a rapporteur plus sign-off from at least a third of the panel, and even then the AI Office can just refuse it. This isn't a rogue AI police force; it's a deliberately slow-moving advisory body wrapped in transparency requirements, publishing its opinions and hearing records while shielding genuinely confidential business data. Whether that balance holds up once a real systemic-risk fight breaks out is the open question nobody can answer yet.
My take — AI-written commentary, not fact-checked reporting
I'll believe this panel has teeth when it actually issues its first qualified alert against a frontier lab, not before. Sixty unpaid-ish experts with a one-third-majority threshold just to ask a company for documents sounds like exactly the kind of committee structure that lets everyone feel like something's being regulated while nothing changes. Europe loves building oversight architecture; the real test is whether the AI Office backs these experts when a US lab pushes back hard.
Read more about this at: AI Act