Art.92
EU AI Act Guide › Chapter IX — Post-Market Monitoring › Article 92

Article 92 – Power to Conduct Evaluations

GPAI Models SME Relevant ~2 min read · 447 words

Article 92 grants the AI Office the power to conduct evaluations of GPAI models — including testing the models directly. This is the most intensive investigative tool available to the AI Office, going beyond requesting documentation to actively assessing whether a model’s capabilities and risks match what the provider claims. It can be triggered by the AI Office itself, the Scientific Panel, or national authorities.

✓ Low compliance impact for SMEs

WHAT THE ARTICLE IS ABOUT

The AI Office’s power to directly test and evaluate GPAI models

Article 92 establishes the AI Office’s power to conduct active evaluations of GPAI models — including running the models, testing their capabilities, assessing their risks, and examining their technical properties. It is the most powerful investigative tool in the GPAI enforcement arsenal and is used when documentation review under Article 91 is insufficient.

WHAT IT SAYS

Evaluations may be triggered on multiple grounds and must involve the provider

  • The AI Office may conduct evaluations of GPAI models to assess compliance with Chapter V obligations where the information gathered under Article 91 is insufficient to draw conclusions
  • Evaluations may also be conducted to investigate systemic risks — particularly following qualified alerts from the Scientific Panel or requests from national market surveillance authorities
  • Providers must cooperate with evaluations — providing access to the model, its infrastructure, and personnel with technical knowledge of the model
  • Evaluations may be conducted by the AI Office itself or by independent experts designated for the purpose
  • The AI Office must respect the confidentiality of commercially sensitive information accessed during evaluations
  • Where an evaluation reveals non-compliance or systemic risks, the AI Office may proceed to request measures under Article 93
  • The costs of evaluations are borne by the AI Office unless the evaluation reveals deliberate non-compliance, in which case costs may be recovered from the provider

WHO IS AFFECTED

GPAI model providers subject to active model assessment

  • GPAI model providers whose models are evaluated — they must cooperate and provide access
  • The AI Office which conducts evaluations or designates experts to do so
  • The Scientific Panel which may trigger evaluations through qualified alerts
  • National market surveillance authorities which may request evaluations

WHAT IT MEANS FOR SMES

Indirect impact through the safety assurance it provides on GPAI models

  • Most SMEs will never be directly subject to Article 92 evaluations — they are aimed at frontier GPAI providers
  • The practical effect is positive: active evaluation of GPAI models by the AI Office means the models SMEs rely on are subject to more rigorous independent assessment than provider self-reporting alone
  • For SMEs building products on GPAI models: an evaluation finding that affects the underlying model may require you to update your product or reassess your AI strategy — monitor AI Office evaluation outcomes
  • The cost recovery provision is a meaningful deterrent against deliberate non-compliance by GPAI providers — where providers are found to have deliberately misled the AI Office, they bear the evaluation costs

Related Articles

← Back to the Guide
← Previous Art. 91 — Power to Request Documentation and Information Next → Art. 93 — Power to Request Measures