The term comes from safety-critical industries such as aviation and nuclear power. For AI, a safety case connects claims about alignment training, containment, monitoring, and operations to the evidence that those claims hold. OpenAI's September 2026 training playbook treats aviation-grade cases as a north star for frontier reinforcement learning, while noting that emergent capability jumps make the same rigor harder than in those industries.