The article discusses the global challenge in the regulation of artificial intelligence (AI), which consists in effective risk management without inhibiting innovation and technical progress[1]. The European AI Act (EU AI Act) is the first binding risk framework in the world that addresses this tension[1]. From August 2025, obligations apply to providers of universal AI models (GPAI), including a systemic risk assessment for the most advanced models[1]. The regulator must ensure that assessments provide meaningful risk information without overburdening providers[1]. The principle of proportionality, binding in EU law, requires adaptation of measures to their objectives[1]. The application of proportionality to the evaluations of AI models makes it possible to develop scientific methods for its practical implementation in evaluations[1]. The article was published in Science, Volume 391, Number 6787, Pages 769-771[1].