Abstract
AI development is driven by performance. Intelligence is judged like art. AI safety, however, has no excuse not to be science. In this keynote, I argue that our reliance on benchmarks, leaderboards, and aggregate metrics is not just limiting – it is misleading. Systems that perform best under current evaluations are not necessarily those we understand, nor those we can trust in real-world settings. I advocate for a shift away from competitive benchmarking toward evaluation practices that prioritize scientific insight. Instead of asking which model scores highest, we should ask how AI systems behave with respect to desired properties, where and why they fail, and what those failures imply for safety, accountability, and trust. By rethinking evaluation, we can move beyond optimizing performance and toward building AI systems that are reliable, interpretable, and worthy of deployment.
