Introduction to AI Safety

Catalog description

What makes an AI system "safe" and how do we decide? CS120 explores this question, focusing on both measurement theoretic and institutional challenges of understanding, interpreting, and governing AI capabilities. We distinguish between AI-specific and systemic safety issues, from examining sycophancy and security implications to the challenge of measuring capabilities and risk. This class will look at current solutions and their limitations through CS publications, and will connect the institutions behind AI, how evaluation is defined and audited, and how those factors shape the future risks we might face. Topics will span data work, foundation models, and governance, focusing on threat-models, evaluation, and auditability. This course aims to prepare you to critically assess and contribute to safe AI development, equipping you with knowledge of cutting-edge research and ongoing debates in the field. This course has no official requirements, although we recommend some knowledge about machine learning and statistics, and will include readings, quizzes, and a final project. For more details, see also the course website: https://stanford-cs120.github.io/fall2026/