Wiki · Governance · Last reviewed June 25, 2026

AI Act Scientific Panel

The AI Act Scientific Panel is the European Union's independent expert body for technical advice on general-purpose AI models, systemic risk alerts, evaluation methods, and market-surveillance support under the EU AI Act.

Category: Governance Published: June 25, 2026 Modified: June 25, 2026 Last reviewed: June 25, 2026 Tags: EU AI Act, AI Office, GPAI, systemic risk, evaluations

Definition

The AI Act Scientific Panel is an expert advisory body established by the European Commission under Article 68 of the EU AI Act and Commission Implementing Regulation (EU) 2025/454. It supports the European AI Office and national authorities on implementation and enforcement, with particular focus on general-purpose AI models and systems.

As reviewed on June 25, 2026, the Commission's public Scientific Panel page said 60 independent AI experts had been recruited. The Commission's June 1, 2026 announcement described the panel and the Advisory Forum as two bodies appointed to support AI Act enforcement, with members serving two-year terms. The Scientific Panel is narrower than the Advisory Forum: it is a technical expert body, while the forum is a broader stakeholder body drawing from industry, civil society, academia, SMEs, startups, and EU agencies.

Mandate

Article 68 gives the panel a concrete enforcement-adjacent role. It advises and supports the AI Office on general-purpose AI models and systems, including systemic-risk alerts, evaluation tools and methodologies, benchmarks, classification of general-purpose AI models with systemic risk, classification of other GPAI models and systems, and development of tools and templates.

The panel can also support market surveillance authorities at their request, support cross-border market surveillance, and assist the AI Office in Union safeguard procedures. That makes it more than a symbolic science council. Its work can shape the technical vocabulary regulators use when asking whether a model is powerful enough, risky enough, or poorly documented enough to trigger further scrutiny.

Systemic-Risk Alerts

Article 90 is the panel's sharpest tool. It allows the panel to provide a qualified alert to the AI Office when it has reason to suspect that a general-purpose AI model poses a concrete identifiable Union-level risk or meets the Article 51 conditions for a model with systemic risk. The alert must be reasoned and include provider contact information, relevant facts, the panel's reasons, and other information the panel considers relevant.

A qualified alert is not a penalty or a public finding that a provider has broken the law. It is a procedural trigger. After such an alert, the Commission, through the AI Office and after informing the AI Board, may use the GPAI supervision powers in the AI Act section that follows Article 90, including requests for documentation, evaluations, and measures where the legal conditions are met.

Independence and Membership

Article 68 requires scientific or technical expertise, independence from providers of AI systems or GPAI models, and the ability to work diligently, accurately, and objectively. The Commission page says experts are appointed in a personal capacity for renewable 24-month terms, with declarations of confidentiality and commitment signed before taking on panel work.

The Commission says selection considered expertise as well as geographic and gender balance. Its public page lists members alphabetically and provides a compiled information file with biographies and declarations of interest. Those disclosures matter because the panel's credibility depends on independence being managed as a continuing condition, not as a one-time appointment label.

Governance Use

For model governance, the Scientific Panel is a review signal to watch when evaluating claims about frontier or widely integrated models. A serious record should ask whether the model was classified as GPAI, whether systemic-risk notification occurred, whether evaluations and benchmarks were documented, whether the provider responded to AI Office information requests, and whether any panel alert or expert advice shaped the classification or mitigation path.

For public accountability, the important artifact is not only the name of the panel. It is the chain from evidence to action: model version, risk claim, benchmark or evaluation method, confidential material, alert or advice, AI Office decision, provider response, and any corrective measure or public explanation.

Limits

The panel is not a court, auditor, notified body, safety institute, standards organization, or replacement for market surveillance authorities. It advises and supports. It does not make a model safe by existing, and panel membership should not be used as a certification badge for any private system.

There is also a transparency tension. Some useful evidence about model capabilities, cybersecurity, or misuse may be confidential. That can be legitimate, but the public record should still identify the kind of decision being made, the authority involved, the model boundary, and which parts of the evidence are restricted rather than absent.

Source Discipline

Use Commission pages for membership, appointment, and institutional role; use the AI Act Service Desk or EUR-Lex for Article 68 and Article 90 text; use Implementing Regulation (EU) 2025/454 for operational rules. Commentary about the panel can explain stakes, but it should not replace dates, article numbers, membership records, declarations of interest, or official AI Office procedure.

Spiralist Reading

Spiralism reads the Scientific Panel as an attempt to put expert friction inside the model frontier. The risk is that the panel becomes a ceremonial witness to decisions already made elsewhere. The promise is narrower and more useful: a public institution with enough technical competence to ask model providers for evidence before capability claims harden into infrastructure.

Sources

Return to Wiki