Measuring AI Safety: Capabilities, Propensities, and Control

As artificial intelligence models grow more advanced, ensuring their safety requires rigorous, quantitative evaluation. Understanding how to measure what an AI can do versus what it tends to do is crucial for responsible development and deployment. This text-based course guides you through the foundational frameworks of AI safety measurement. You will transition from a basic understanding of AI risk to practically evaluating model capabilities, assessing behavioral propensities, and testing safety control mechanisms. What you'll learn: 1. Understand the core concepts of AI safety evaluation, including threat modeling and risk taxonomy. 2. Measure model capabilities to identify the maximum potential risks and boundaries of advanced systems. 3. Analyze behavioral propensities to predict how models act in open-ended or adversarial environments. 4. Evaluate control effectiveness by testing safety guardrails, alignment techniques, and system interventions. 5. Apply modern red-teaming concepts and automated evaluation frameworks to real-world scenarios. 6. Practice designing safety test suites through structured written exercises and case studies. The course begins with foundational definitions of AI safety metrics before moving into practical methodologies for measuring capabilities and behaviors. You will explore how to analyze control systems and implement modern evaluation standards through clear, written explanations. This course is designed for beginners, developers, and policy enthusiasts who want to understand AI safety auditing, with no advanced technical prerequisites required. Start reading today to build a strong foundation in modern AI safety measurement and risk assessment.