Run adversarial model evaluation and report serious: Beyond the duties owed by every GPAI provider, which

Adversarial model evaluation and serious-incident reporting to the AI Office are additional obligations that apply specifically to systemic-risk GPAI models.

The question

Preparing for AIGP? Take the free 5-min readiness quiz →

1. Publish a sufficiently detailed summary of the content used to train the model.

   Plausible but wrong: publishing a training-data summary is a baseline duty for all GPAI providers, not a systemic-risk-only obligation.
2. Put a copyright policy in place covering text-and-data-mining reservations by holders.

   Plausible but wrong: adopting a copyright policy is part of the GPAI baseline that applies regardless of systemic risk.
3. Maintain technical documentation and share it with all downstream model providers.

   Plausible but wrong: keeping technical documentation and sharing it downstream is a baseline GPAI duty, not specific to systemic-risk models.
4. Run adversarial model evaluation and report serious incidents to the AI Office. ✓

   Correct: systemic-risk GPAI providers must perform state-of-the-art model evaluation including adversarial testing, mitigate systemic risks, ensure cybersecurity, and report serious incidents to the AI Office.

How to remember it

Adversarial model evaluation and serious-incident reporting to the AI Office are additional obligations that apply specifically to systemic-risk GPAI models.

More Understanding How Laws, Standards and Frameworks Apply to AI questions

• Human oversight: Which EU AI Act requirement does this describe? →
• A cap set at a fixed euro sum or a share of global annual: How are administrative fines generally structured →
• The non-EU company is the provider: Under the EU AI Act, what roles do the two companies hold respectively? →
• All 394 Understanding How Laws, Standards and Frameworks Apply to AI questions →