The science and practice of proportionality in AI risk evaluations

A major international research team of 22 authors, led by Carlos Mougan, has published a foundational paper titled 'The science and practice of proportionality in AI risk evaluations.' The work directly addresses a core tension in the world's first comprehensive AI regulation, the European Union's AI Act. With obligations for providers of advanced General-Purpose AI (GPAI) models like GPT-4o and Claude 3.5 now in effect, these companies must evaluate potential systemic risks. The paper argues that to be effective and legally sound, these mandatory evaluations must adhere to the EU's binding legal principle of proportionality, meaning regulatory actions must be calibrated to their objectives.

The researchers propose that this legal requirement opens a critical opportunity for computer science. Instead of vague guidelines, they advocate for developing concrete, scientific methods to operationalize proportionality within AI risk evaluation practices. This involves creating frameworks to determine the appropriate scale, scope, and rigor of testing needed for different models based on their capabilities and potential impact. The goal is to prevent a one-size-fits-all approach that could either stifle innovation with excessive overhead or fail to catch genuine risks with insufficient scrutiny.

Published in the journal *Science* and on arXiv, this paper is positioned as a guide for both regulators and AI labs like OpenAI, Anthropic, and Google DeepMind. It provides a scholarly foundation for designing evaluation regimes that are both rigorous and efficient. As global AI governance takes shape, this research highlights the need to embed legal principles like proportionality directly into the technical practices of AI safety and alignment, moving the debate from abstract policy to actionable engineering standards.