GPAI model with systemic risk

Some general-purpose models are so capable that their impact could affect society as a whole. The AI Act calls them GPAI models with systemic risk. A model is presumed to be one when the compute used to train it exceeds 10²⁵ FLOPs, although the Commission can also designate one on other criteria.

On top of the general GPAI obligations, their providers must evaluate the model with adversarial testing, assess and mitigate systemic risks, report serious incidents to the AI Office and ensure adequate cybersecurity.

In practice this affects very few companies worldwide: the large AI labs. For a small business, what matters is knowing whether the model it uses falls in this category, because that signals its provider is under stricter scrutiny.