EU AI Office publishes requirements for GPAI evaluators

The European Commission's AI Office has set out what an outside organisation must look like to evaluate a general-purpose AI model with systemic risk. In a report published on 21 October 2026, it says evaluators need "domain expertise and the necessary evaluation skills for systemic risk assessment" and must be "free from the provider's managerial, financial or operational influence."

The document is the follow-up to a 15 July workshop that brought together experts from 19 organisations, selected from twenty invited through a public call. The AI Office is building a system for scoring candidate evaluators against those two axes, qualification and independence, with the workshop specifically addressing "criteria for scoring financial and management independence."

The legal anchor is Article 55(1)(b) of the AI Act, as specified in the GPAI Code of Practice, which make external evaluation a required input to assessing and mitigating systemic risks from frontier models. The AI Office says its written report also addressed questions and concerns raised by the workshop participants.