The GPAI Systemic Risk Manager is the management owner of systemic-risk obligations for general-purpose AI models with systemic risk (classification, Art. 51; obligations, Art. 55) that the organisation provides, and of systemic-risk due diligence for those it integrates. Role 1.4 owns the enterprise risk process. Role 3.2 owns security. Role 8.6 owns the authority clock. This role makes sure model evaluations, adversarial testing, serious-incident tracking and cybersecurity protection for GPAI models with systemic risk exist, are current and feed the risk file.

Standing duties

• Establish systemic-risk rules for GPAI models: classification triggers (high-impact capabilities, Art. 51(1)(a), presumed when training compute exceeds 10²⁵ floating-point operations, Art. 51(2); or Commission designation against Annex XIII, Art. 51(1)(b)); notification to the Commission within two weeks (Art. 52(1)); model evaluation and adversarial testing (Art. 55(1)(a)); assessment and mitigation of systemic risks (Art. 55(1)(b)); serious-incident tracking (Art. 55(1)(c)); and cybersecurity protection (Art. 55(1)(d)).
• Recommend at Implement to Role 15.1 (release gate): no GPAI model with systemic risk is placed or integrated into high-risk agentic systems without current evaluations and mitigations (no veto; ROLES13, t269).
• Track serious incidents and possible corrective measures for GPAI models with systemic risk and make sure they are reported without undue delay to the AI Office and, as appropriate, national competent authorities (Art. 55(1)(c)), through Role 22.1 and Role 8.6.
• Raise ACT02 when model capabilities or integration change the systemic-risk profile; raise ACT03 on serious incidents or evaluation results outside appetite.
• Do not treat a model card, benchmark score or provider statement as a systemic-risk evaluation or as residual-risk acceptance.

Accountable to: AI Governance Lead / Head of AI Risk Management (or GM: Compliance)

Primary domain: Establish (systemic-risk rules for general-purpose AI models) and Review (model evaluation evidence). Does not Direct or accept residual risk.

Linked EN 18286 clauses: 7.1, 8.4.1, 8.4.2, 8.4.3.1, 8.4.3.2, 9.3.1

Format: 3 hours of training, EUR 200.