The Act places models that could affect millions of users or critical systems at scale in a heightened category above standard GPAI. The current trigger is cumulative training compute exceeding 10²⁵ FLOPs; the Commission can change it. section 7.2.4
Providers of systemic-risk models have additional duties:
- Conduct adversarial testing and systemic-risk assessments, report incidents to the AI Office, and apply stronger cybersecurity controls. section 7.2.4
- If your system uses one of these models, it does not become systemic-risk solely for that reason; ask the provider for documentation, safety reports, and incident notices during vendor-risk review. section 7.2.4
