A framework to categorise modified general-purpose AI models as new models based on behavioural changes : collection of external scientific studies on general-purpose AI models under the EU AI Act

This report discusses an approach for measuring behavioural change in models following modification, which could be used as part of a broader assessment to determine when a modified General-Purpose AI (GPAI) model should be considered a new and distinct model for regulatory purposes under the EU AI Act. It presents two approaches to assess behavioural changes in altered models: (1) directly measuring differences in capability profiles or instance-level answers, and (2) using proxy metrics related to the alteration process, such as finetuning, computation, and data usage. The report highlights the challenges of establishing thresholds for determining when an altered model might be considered a new one based on behavioural change, and suggests empirical studies to validate the relationships between alteration metrics and downstream behavioural changes.