Skip to main content
Knowledge4Policy
Knowledge for policy
Supporting policy with scientific evidence

We mobilise people and resources to create, curate, make sense of and use knowledge to inform policymaking across Europe.

  • News | 19 Feb 2026
New paper in Science: The science and practice of proportionality in AI risk evaluations

A new paper co-authored by the AI Office and the JRC has been published in Science.

AI Watch
Headline in Science publication

The paper explores an interdisciplinary research question: how can AI evaluations provide meaningful information about risks without imposing excessive burden? 

As AI models become more capable and widely used, providers of the most advanced general-purpose AI (GPAI) models are required to assess and mitigate possible systemic risks under the EU AI Act. But deciding how much testing is enough, and how demanding those tests should be, is not straightforward.  

Finding the right balance 

The paper explains how the principle of proportionality, a core principle of EU law, can help strike the right balance. It translates the three legal steps of proportionality (suitability, necessity, and balancing) into concrete criteria for AI evaluations: 

  • Suitability examines whether an evaluation provides sufficient informational value about a specific risk.
  • Necessity requires comparing evaluations with comparable informational value to determine whether less intrusive or resource demanding alternatives are available.
  • Balancing weighs the informational value of an evaluation against its burden. 

In practice, this means that researchers developing AI evaluations should document and compare the effectiveness and burden of their evaluations with those of other relevant evaluations, to facilitate proportionality assessments. 

This paper contributes to the growing global debate on how to achieve effective risk management without compromising innovation and technical progress. 

GPAI eval paper - Fig1
Proportionate evaluations of AI models involve a trade-off between how effective and how burdensome evaluations are. Some evaluations provide better insight into potential risks but require more time, data, computing power, or access to sensitive systems. Each curve in the figure represents a different type of evaluation approach, and each point shows a specific version of that approach (for example, using more samples or more advanced testing techniques)

A collaborative effort 

This publication reflects a joint effort by researchers from the EU AI Office and the Joint Research Centre (JRC), building on interdisciplinary collaboration developed following a workshop organised by the AI Office in April 2025. 

Congratulations to the authors from the EU AI Office: Carlos Mougan, Lauritz Morlock, Jan Brauner, Alberto Franzin, Friederike Grosse-Holz, Eloise Hamilton, Max Hasin, Luca Massarelli, and Wout Schellaert, and to the authors from the JRC: David Fernández Llorca and Emilia Gómez. 

This paper can now be cited as: 

Mougan, C., Morlock, L., et al. The science and practice of proportionality in AI risk evaluations. Science (2026) 

Read the paper here

Disclaimer 

The views expressed in this publication do not necessarily reflect those of the European Commission and are made in a personal capacity. The European Commission and any person acting on behalf of it are not responsible for the use that might be made of this publication. 

While the paper does not represent the views of, nor commit, the AI Office, it is a significant contribution made by its staff to the emerging science and practice of AI evaluations.