# AI Safety as an Empirical Science **Entity class:** Research-program concept **Domain:** Artificial intelligence / safety / scientific method **Maturity:** Developed ## Definition **AI safety as an empirical science** treats safety claims as questions to be tested through models, experiments, evaluations, reproducible methods, peer review, incident evidence, and adversarial investigation rather than resolved only through intuition or abstract argument. ## Research program The empirical program includes alignment experiments, interpretability, reward-hacking analysis, capability evaluations, security testing, behavioral monitoring, and study of failure under distribution shift. It connects [[wiki/AI Safety|AI Safety]], [[wiki/Scientific Literacy|Scientific Literacy]], [[wiki/AI Interpretability|AI Interpretability]], and [[wiki/Reward Hacking|Reward Hacking]]. ## Debate boundary Empirical work does not remove philosophy or governance from AI safety. It constrains them by separating demonstrated behavior, model-specific evidence, plausible mechanisms, and speculative scenarios. ## Source route - [[research/Vishal Maini - Humanity's Machine Successor and the AI Transition|Humanity's Machine Successor]] ## Evidence boundary A published result is not automatically general across architectures, scales, deployments, or future systems. Empirical safety requires repeated measurement and explicit limits on inference. ## Simple Reminders, Quotations, and Thoughts > "The people who internalized that human-level AI might be around the corner were more willing to spend real effort, research energy, and compute on making AI safety an empirical science—not something confined to blog posts or philosophical essays, but a field grounded in machine-learning research, real conference papers, peer review, and rigorous scientific and academic debate about the properties of these systems. The field grew faster because some people were willing to bet on its importance very early." > **— Vishal Maini**, *Palisade Research interview, September 29, 2026* [[reminders/Scientific Literacy/AI Safety Must Become an Empirical Science by Vishal Maini|AI Safety Must Become an Empirical Science by Vishal Maini]]