Home/Recitals/Recital 99
Recital (99)

Systemic Risk — Definition and Classification

Plain-Language Summary

Defines systemic risk for GPAI models and explains the 10^25 FLOPs threshold as a proxy for high-impact capability sufficient to justify additional obligations.

Keywords

systemic riskGPAI10^25 FLOPsclassificationhigh-impact capabilities

Legal Text

General-purpose AI models with systemic risk are those having high impact capabilities which, due to their reach, can have a significant negative effect on public safety, security or fundamental rights across multiple Member States. The threshold of 10^25 floating-point operations for training is used as a proxy indicator of high-impact capabilities. This threshold may be updated by the Commission as technology evolves. Models meeting the threshold are presumed to have systemic risk, though providers may rebut this presumption.
Recitals are the preamble paragraphs of the EU AI Act. They explain the intent behind provisions but are not directly binding.