Systemic Risk: where individual model risk becomes market-wide risk
Systemic AI risk affects multiple deployers via shared models. Zertia audits GPAI under ISO 42001 + EU AI Act Articles 53-55.
Definition
Systemic AI risk affects multiple deployers via shared models. Zertia audits GPAI under ISO 42001 + EU AI Act Articles 53-55.
Key facts
- Systemic risk in AI refers to risks affecting multiple systems, sectors or population groups simultaneously due to shared infrastructure or correlated failures.
- EU AI Act Article 51 specifically designates GPAI with systemic risk above 10^25 FLOPs training compute.
- Systemic-risk GPAI faces Article 55 obligations including adversarial evaluation, incident reporting and cybersecurity measures.
- Systemic AI risks include market-wide hallucination patterns, correlated failures across deployers, and cascading effects across dependent systems.
- Zertia audits systemic-risk GPAI providers under ISO 42001 with explicit Articles 53-55 mapping.
Frequently asked questions
What is systemic risk in AI?
Risk affecting multiple systems, sectors or population groups simultaneously, typically arising from shared model infrastructure (e.g., the same foundation model deployed by thousands of organizations) where individual incidents become market-wide failures.
How does EU AI Act define systemic risk?
Article 51 designates GPAI with systemic risk presumptively when cumulative training compute exceeds 10^25 FLOPs. These models face Article 55 obligations: adversarial evaluation, incident reporting, cybersecurity measures and post-market monitoring.
How is systemic-risk GPAI audited?
Through ISO 42001 with explicit Articles 53-55 mapping. Zertia provides ANAB-accredited audit for GPAI providers.
Definitions that hold up under audit.
Does this term apply to your certification project? Let's talk 30 minutes, no commercial pressure.