New paper in Science: The science and practice of proportionality in AI risk evaluations

2026-02-19 · Source: AI Watch | News · Field: Legal & Regulatory — Compliance & Risk Management, Regulatory Affairs & Government Relations · Depth: Advanced, quick

Summary

A new paper, "The science and practice of proportionality in AI risk evaluations," published in Science (2026) by researchers from the EU AI Office and the Joint Research Centre (JRC), addresses the challenge of balancing effective AI risk evaluations with avoiding excessive burden. The study proposes applying the EU legal principle of proportionality to General-Purpose AI (GPAI) model assessments, which are mandated by the EU AI Act. This framework translates proportionality into three criteria for AI evaluations: suitability (informational value for risk), necessity (comparing less intrusive alternatives), and balancing (weighing informational value against burden). The paper emphasizes documenting and comparing evaluation effectiveness and burden to facilitate these assessments, contributing to the global discourse on AI risk management and innovation.

Key takeaway

For AI Scientists and Research Scientists developing or implementing risk assessments for GPAI models, you should integrate the principle of proportionality into your evaluation design. Documenting the informational value and resource burden of your chosen evaluation methods, and comparing them against less demanding alternatives, will help ensure compliance with regulations like the EU AI Act while fostering innovation.

Key insights

Proportionality, a core EU legal principle, can guide effective and efficient AI risk evaluations.

Principles

• Evaluations must provide sufficient informational value.
• Less intrusive alternatives should be prioritized.
• Informational value must outweigh evaluation burden.

Method

Researchers developing AI evaluations should document and compare the effectiveness and burden of their evaluations against relevant alternatives to facilitate proportionality assessments.

In practice

• Apply suitability criteria to assess evaluation informational value.
• Use necessity to compare evaluation intrusiveness.
• Balance evaluation value against its resource demands.

Topics

• AI Risk Evaluation
• EU AI Act
• General-Purpose AI
• Proportionality Principle
• AI Governance

Best for: AI Scientist, Research Scientist, CTO, AI Researcher, Policy Maker, AI Ethicist

Related on AIssential

• Charting the AI Perception Gap: Across 71 scenarios, AI experts (N=119) and the public (N=1100) have differing views on the risks, benefits, and value of AI. More importantly, AI experts discount the influence of risks stronger than the public does when forming their value judgments [R] — AI experts and the public hold systematically different mental models regarding AI's risks, benefits, and overall value.
• The EU’s Real AI Leverage Is Making Compliance the Path of Least Resistance — Interoperable AI compliance standards, driven by economic incentives, can make EU regulations a global baseline.
• Import AI 446: Nuclear LLMs; China's big AI benchmark; measurement and AI policy — Effective AI governance and safety require robust measurement, while LLMs exhibit aggressive tendencies in conflict simulations.
• AI Fairness Metrics Explained: A Practical Guide with Python — Measuring AI fairness requires selecting appropriate metrics as different definitions often conflict.
• Mathematicians warn of AI threats to profession as industry encroaches — AI's rapid advancement threatens the integrity, autonomy, and human-centric values of mathematical research and its institutions.
• A Framework for Measuring Appropriate Reliance on Set-Valued AI Advice — This framework is the first to formally measure appropriate reliance on set-valued AI advice in human-AI collaboration.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI Watch | News.