Article 51 establishes a two-tier classification for GPAI models: standard GPAI models and GPAI models with systemic risk. The systemic risk designation triggers a significantly enhanced obligation set under Article 55.
The 10^25 FLOPs Threshold
Article 51(2) creates a presumption of systemic risk for GPAI models trained using a total computing power of more than 10^25 floating-point operations (FLOPs). This threshold was chosen as a proxy for capability level because training compute strongly correlates with model capability in current large-scale AI systems. At the time of the regulation's adoption, models such as GPT-4 and Gemini 1.0 Ultra were understood to be in this range.
AI Office Designation
Article 51(1) also allows the AI Office to designate additional GPAI models as presenting systemic risk based on qualitative criteria in Article 51(1)(a) -- "high impact capabilities evaluated on the basis of appropriate technical tools and methodologies, including indicators and benchmarks." The AI Office may issue a qualified alert through the scientific panel. Providers may rebut this designation by demonstrating that the model does not present systemic risk.
Regulation (EU) 2024/1689 — Article 51(2)
“A general-purpose AI model shall be presumed to have high impact capabilities pursuant to paragraph 1, point (a), when the cumulative amount of compute used for its training measured in floating point operations (FLOPs) is greater than 10^25.”
- The 10^25 FLOPs threshold creates an automatic presumption of systemic risk -- notification to the AI Office is required
- Providers of models below the threshold may still be designated systemic risk based on qualitative criteria
- The AI Office publishes guidance on computing metrics and assessment methodologies
- Systemic risk models must comply with both Article 53 baseline obligations AND Article 55 enhanced obligations
- The Commission may update the compute threshold by delegated act as technology evolves