The Science and Practice of Proportionality in AI Risk Evaluations

A global challenge in artificial intelligence (AI) regulation lies in achieving effective risk management without compromising innovation and technical progress. The European Union (EU) Artificial Intelligence Act represents the first regulatory attempt worldwide to navigate this tension in the form of a binding, risk-based framework. In August 2025, obligations for providers of general-purpose AI (GPAI) models under the EU AI Act entered into application. They require providers of the most advanced GPAI models to evaluate possible systemic risks stemming from their models. This raises the regulatory challenge of ensuring that the evaluations provide meaningful risk information without imposing excessive burden on providers. The principle of proportionality, a binding requirement under EU law, requires the regulator to calibrate its actions to their intended objectives. The application of proportionality to model evaluations for AI risk opens opportunities to develop scientific methods that operationalize such calibration within concrete evaluation practices.

According to the principle of proportionality, EU measures must be suitable, necessary, and balanced. A measure is suitable if it pursues an objective that is aligned with the intent of its legislative basis and is implemented consistently and systematically; in other words, it must meet a minimum level of effectiveness for a legitimate aim. It is considered necessary for the level of effectiveness that it achieves if there is no less restrictive yet equally effective means to achieve its objective; in other words, it cannot be a case of a sledgehammer cracking a nut. It is balanced if its burden does not clearly outweigh its benefit. in other words, even if a measure is suitable and necessary, it must not appear excessive. Regulators have some discretion in determining whether a measure meets these conditions, if doing so requires a complex and evidence-based case-by-case assessment that a court would not be better placed to replicate in hindsight.

Recent work emphasizes the importance of evidence-based AI policy and the corresponding need to accelerate evidence generation. Nevertheless, it does not address the regulatory question of how much evaluation effort is legally justifiable, i.e., of what proportionality means in practice. Although this question arises in other regulatory domains as well, applying it to AI model evaluations poses distinct challenges. In the absence of methodologies to determine when requiring an evaluation is proportionate, regulators face limited guidance in navigating complex trade-offs, expanding the scope for normative judgment and increasing the risk of regulatory outcomes that are insufficiently protective or unduly burdensome.