High-impact / systemic risk (AI Regulation)
Art. 51 AI Act — classification of high-capability GPAI
TL;DR
The systemic risk classification (Art. 51 AI Act) applies to GPAI models of particularly high capability. Threshold: cumulative training compute operations >10^25 FLOPs.
What is high-impact / systemic risk (AI Act)?
As at 04/2026: around 25 systemic-risk models worldwide. Examples: GPT-4o, Claude 3 Opus/3.5 Sonnet, Gemini 1.5/2 Pro, Llama 3.1 405B, Mistral Large 2.
Practical example
OpenAI GPT-4 with around 10^26 FLOPs of training: systemic risk. Obligations: adversarial testing, risk assessment, reporting to the AI Office.
Frequently asked questions
Who decides?
The EU AI Office, on the basis of provider information + technical assessment.
Does it affect SMEs?
No, almost never. Only providers of such models (OpenAI, Anthropic, etc.).
See also
Practitioner note: This article is practice-oriented compliance documentation, not legal advice. We are compliance specialists, not a law firm. For legally binding advice, please consult an admitted lawyer.