The AI Act's expert bodies: the Scientific Panel and the Advisory Forum

The AI Act is enforced, but until mid-2026 the people who would do the technical thinking behind that enforcement did not yet exist as a body. On 1 June 2026 the European Commission closed that gap, appointing the two expert bodies the regulation itself foresees: the Scientific Panel of independent experts and the Advisory Forum. Neither writes rules. Both exist to make sure the rules that the AI Office and national authorities apply rest on current science and broad input rather than on the Commission's own reading alone.

The Scientific Panel: 60 experts, no instructions

The Scientific Panel is created by Article 68 of the AI Act, fleshed out by Commission Implementing Regulation (EU) 2025/454. The Commission recruited 60 independent experts with standing in frontier AI, engineering, technical auditing, industry and societal impact. They serve a renewable 24-month term in a personal capacity — they "neither seek nor take instructions from anyone", must declare their interests publicly, and must be independent of any provider of AI systems or general-purpose AI models. Selection balanced geography (at most three nationals of any one country, at least 80% from EU/EFTA/EEA states) and gender.

The panel's centre of gravity is general-purpose AI. Its tasks are to:

• alert the AI Office to systemic risks of GPAI models;
• advise on the classification of GPAI models with systemic risk and on

evaluation methodologies, tools and benchmarks;

• support market surveillance authorities, at their request, including in

cross-border activities.

That last task matters: a national market surveillance authority confronting a frontier model usually lacks the in-house capacity to evaluate it. The panel is the shared technical reserve they can draw on — the mechanism that keeps enforcement of the GPAI regime from depending on whichever member state happens to host the provider.

The Advisory Forum: 174 voices, standards built in

The Advisory Forum is created by Article 67. Where the panel is small and technical, the forum is large and representative: 174 members, selected from more than 700 applications, drawn from academia, civil society and industry — expressly including SMEs and startups. They too serve two-year renewable terms (capped at two terms) and advise both the Commission and the AI Board of member states.

Five organisations sit on the forum as permanent members: the EU Fundamental Rights Agency, the cybersecurity agency ENISA, and the three European standardisation bodies CEN, CENELEC and ETSI. Their presence is deliberate. The forum's remit explicitly covers standardisation and implementation challenges, and the harmonised standards under Article 40 — the practical route to a presumption of conformity for high-risk systems — are exactly where the AI Act's timetable has slipped. Putting the standards bodies in the room is an acknowledgement that compliance lives or dies on standards that are still being drafted.

Where the two bodies sit in the enforcement architecture

It helps to keep four institutions apart:

• The AI Office (inside the Commission) enforces the GPAI rules and

coordinates the system.

• The AI Board brings the member states together.
• The Scientific Panel supplies independent technical expertise, weighted to

GPAI and systemic risk.

• The Advisory Forum supplies broad stakeholder and standardisation input.

The panel and the forum are advisory: they do not impose obligations, levy fines or take binding decisions. But their appointment is not cosmetic. It lands deliberately ahead of 2 August 2026, the date from which the AI Office can sanction GPAI providers and the next wave of obligations begins. From that point the Commission needs a standing source of evidence it cannot be accused of authoring itself — for classifying systemic-risk models, for judging whether the GPAI Code of Practice is doing its job, and for backing national regulators who lack frontier-model expertise.

What it means

The significance of 1 June 2026 is institutional, not substantive: no new obligation arrived. What arrived is the machinery that makes the existing obligations credible. An AI Act enforced on the Commission's unaided judgement would be contestable in every systemic-risk classification; an AI Act enforced with a published, independent panel and a 174-member forum behind it is much harder to dismiss as political. For providers of general-purpose AI in particular, the practical lesson is that the body which will second-guess a "no systemic risk" claim now exists, has names attached, and reports to the regulator — not to them.