General-Purpose AI Model Reach as a Criterion for Systemic Risk
Röttger, P., Hovy, D., Nozza, D., Vanschoren, J., Fernandez Llorca, D., Noroozian, A., Gomez, E.
Röttger, P. et al. - European Commission, Joint Research Centre (Publications Office of the European Union), 2025-10-10
Abstract
Under the EU AI Act, systemic risks of general-purpose AI (GPAI) models are assumed to increase with model "reach". This report proposes a method for operationalising, and measuring "reach" so that it may be used as a criterion when determining whether a GPAI model should be classified as a GPAI model with systemic risks. This report considers specific systemic risks created by model reach, such as risks from model biases, i.e. systemic risk which become relevant when models are widely used, regardless of whether their capabilities are at the technological frontier. In this report, reach is defined as the number of people who directly interact with a given GPAI model. Potential metrics for measuring reach are identified, including access via user interfaces and APIs. A concrete proposal for the design of a reach trigger is presented, which includes overall reporting thresholds for model providers, reach metrics to be reported, and calibration of the reach trigger.