A structured method for constructing safety cases for AI systems

The researchers propose a structured method for constructing safety cases for AI systems, used to argue that the safety risks of generative and frontier AI systems are acceptable. The paper points out that traditional safety case practices in fields such as aviation and nuclear power rely on clear system boundaries, stable architectures, and known failure modes, and struggle to characterize the unpredictable emergence of generative and agentic AI capabilities, behavior that changes with prompts, and risks that drift with fine-tuning, scaffolding, and deployment environments. The study provides a taxonomy of AI-specific claim types (assertional, constraint-based, capability-based), argument types (demonstrative, comparative, causal/explanatory, risk-based, normative), and evidence families (empirical, mechanistic, comparative, expert-driven, formal methods, operational/field data, model-based), and proposes reusable safety case templates, together with end-to-end patterns to address challenges such as evaluations without ground truth, dynamic model updates, and threshold-based risk decisions, making safety cases credible, auditable, and able to adapt to evolving model behavior.