Article 67: Scientific Panel of Independent Experts

Article 67 creates a panel of independent scientific and technical experts whose primary job is to provide rigorous, evidence-based advice to the AI Office. The panel is not a policy or stakeholder body (that role belongs to the advisory forum) — it is an expert resource for enforcement-critical technical questions.

The panel's core tasks include:

1. (a) GPAI classification and systemic risk: advising the AI Office on whether a general-purpose AI model should be classified as presenting systemic risk, including assessing capabilities, reach, and potential harms. 2. (b) Evaluation tools and methodologies: contributing to the development of tools, benchmarks, and methodologies for evaluating GPAI model capabilities — particularly for models with systemic risk. 3. (c) Risk alerts: issuing qualified alerts to the AI Office when a GPAI model may pose specific risks at Union level. 4. (d) Model evaluations: advising on and supporting model evaluations carried out by the AI Office, including evaluations under Article 93.

Independence is paramount: panel members are selected based on up-to-date scientific or technical expertise and must be free from conflicts of interest. The selection criteria and procedures are set out in the article — see EUR-Lex Article 67.

• Article 51 — Classification of GPAI models with systemic risk: the scientific panel advises on whether a model crosses the systemic-risk threshold.
• Article 55 — Obligations for providers of GPAI models with systemic risk: the panel's assessments feed into enforcement of these obligations.
• Article 56 — Codes of practice for GPAI: the panel's technical findings may inform the substance of codes of practice.
• Article 64 — AI Board tasks: the Board may draw on the panel's expertise indirectly via the AI Office.
• Article 68 — Member State access to the pool of experts: national authorities can call on the same experts for national enforcement.
• Article 93 — Compliance and enforcement powers: the panel supports AI Office evaluations under these provisions.
• Article 113 — Application dates: the governance architecture applies on the timeline set out in Article 113.

If you develop or distribute general-purpose AI models:

• Expect panel-informed evaluations: the AI Office may rely on the scientific panel's advice when assessing whether your model has systemic risk under Article 51. Prepare documentation on model capabilities, training data, and compute.
• Monitor panel publications: the panel's work on evaluation methodologies and benchmarks will shape what the AI Office expects in practice. Tracking panel outputs lets you anticipate evaluation criteria.
• Engage constructively with evaluations: if the AI Office initiates a model evaluation (see Article 93), the panel's technical findings will inform the outcome. Transparent cooperation is in your interest.

If you are a national authority, see Article 68 for how you can access the panel's expertise for your own enforcement activities.

Article 67 places significant emphasis on independence. Panel members must:

• Possess up-to-date scientific or technical expertise in AI-relevant disciplines.
• Be free from conflicts of interest — including financial, professional, or institutional ties that could compromise objectivity.
• Comply with confidentiality obligations regarding information obtained in the course of their duties.

These requirements are designed to ensure that the panel's advice carries scientific credibility and cannot be dismissed as captured by any particular interest group. The full selection and governance procedures are set out on EUR-Lex Article 67.