The EU AI Office Scientific Panel Is Operating. What Does Independent Mean Here.

On the first standup of expert structures under the AI Act, and the reporting line that sits underneath the word "independent."

On June 1, 2026, the European Commission announced that it had appointed a Scientific Panel and an Advisory Forum to support enforcement of the Artificial Intelligence Act. This is the first operational standup of the expert structures the AI Act envisioned to support systemic-risk evaluation and market surveillance for general-purpose AI models. The Scientific Panel comprises 60 experts described by the Commission as independent, drawn from frontier AI research, engineering, technical auditing, industry, and societal-impact backgrounds. The Advisory Forum draws from academia, civil society, and industry, including small and medium-sized enterprises and startups, with permanent seats reserved for the EU Agency for Fundamental Rights, the EU Agency for Cybersecurity, and relevant standardisation bodies. Members of both bodies serve two-year terms.

The Commission's own framing names the function precisely: the two bodies will advise the Commission's AI Office and national authorities on applying the Act's rules. The Scientific Panel's mandate covers general-purpose AI models and systems, systemic risk, model classification, evaluation methodologies, and cross-border market surveillance. The Advisory Forum's mandate is broader and more procedural, covering standardisation and implementation questions across the Act's full scope. Neither body is described as an enforcement authority in its own right. Both are described as advisory to one.

The Word Doing the Work

"Independent" appears repeatedly in the Commission's announcement, attached to the panel's composition rather than to its institutional position. The 60 members are independent in the sense that they are not Commission staff and do not hold their seats by virtue of employment inside the AI Office. That is a real form of independence, and it is not a small one: expertise drawn from outside a regulator's own payroll is difficult to manufacture internally and expensive to substitute with in-house staff.

But the panel's output routes through a single channel. It advises the AI Office. It does not publish independent determinations that bind the AI Office, and it does not appear, from the Commission's own description, to have a reporting line that runs anywhere except toward the body it was created to support. This is a structurally different kind of independence than, for instance, an external auditor whose findings a company cannot suppress before they reach a regulator or the public. A scientific panel whose only formal audience is the entity it evaluates on behalf of has independence of composition without independence of publication. Whether that distinction matters in practice depends on details the June 1 announcement does not specify: whether panel findings are published in full, in summary, or not at all when they diverge from the AI Office's own enforcement posture.

Systemic Risk and the Post-Market Question

The Scientific Panel's mandate over "systemic risks" and "cross-border market surveillance" points toward one of the AI Act's more difficult implementation problems: general-purpose AI models are not evaluated once and then left alone. Systemic risk, as the Act's architecture generally treats it, is a property that can emerge or change after deployment, as a model is fine-tuned, scaled, integrated into new downstream applications, or exposed to novel use patterns the original evaluation did not anticipate. A panel with a genuine post-market monitoring function would need visibility into how a general-purpose model is behaving in the field, not only how it performed in a pre-deployment evaluation.

The Commission's announcement does not specify the panel's access to post-market data, the cadence at which it will review deployed systems, or whether providers are obligated to furnish updated information as models change. It also does not specify whether the panel's evaluation methodologies, one of its stated focus areas, will be published in a form that allows outside verification of how a systemic-risk classification was reached. Evaluation methodology developed inside an advisory structure and published only as a summary is a different accountability artifact than a methodology published in full and subject to outside replication.

What Activation Signals, and What It Does Not

The significance of this announcement is that it is the first operational instance of these bodies, not merely their creation on paper. The AI Act named these structures well before June 2026; the news is that the seats are now filled and the bodies are functioning. That distinction is worth taking seriously: a statute that authorizes an advisory body and a Commission that actually appoints sixty people to staff it are not the same level of institutional commitment. The staffing decision is real. It is also, on its own, silent about the harder question of what the panel can do when its independent judgment about a model's systemic risk differs from the AI Office's own enforcement priorities or resource constraints.

The comparison worth drawing is not to a court, which can rule against the party that created it, but to an internal audit committee that reports to the same executive whose decisions it reviews. Internal audit committees can be genuinely rigorous and genuinely independent in composition while still operating inside a reporting structure that gives the audited party control over what becomes visible outside the organization. Whether the Scientific Panel functions closer to the court model or the audit-committee model is not resolved by the June 1 announcement. It will be resolved, if it is resolved at all, by how the panel's findings are handled the first time they conflict with an AI Office enforcement decision.

Open Questions

• At what point does a panel's advisory output become public regardless of whether the AI Office agrees with its conclusions, and who makes that publication decision.
• What access does the Scientific Panel have to post-market performance data for general-purpose AI systems already in deployment, as distinct from pre-market evaluation data.
• How does a national authority distinguish between the AI Office adopting the panel's technical judgment and the AI Office overriding it, when both outcomes could be described publicly in the same language.
• What standard applies when the panel's systemic-risk classification changes after deployment, and who is obligated to act on that changed classification.
• How does an outside party verify that an evaluation methodology developed by the panel was actually applied consistently across the general-purpose models it reviews.
• If a panel member's independent assessment is not adopted by the AI Office, is that disagreement recorded anywhere a regulator, company, or member state can later examine.

The Scientific Panel and Advisory Forum are now staffed and operating, which is more than existed a month earlier. Whether that operational fact converts into a check on the AI Office rather than a resource for it is not something the June 1 announcement settles. The policy instruments and the deployment tempo are not aligned.