A safety case is a structured document making a reasoned, evidence-backed argument that a system is safe enough to operate. The concept comes from aviation and nuclear power, where regulators have required formal written arguments for decades before a plane flies or a reactor goes online.

The argument isn't just "we tested it and it passed." It defines specific hazards, marshals evidence from testing and monitoring, and honestly accounts for residual risks that remain. It's a living document, updated as the system changes or new failure modes emerge.

Organizations like the UK AI Safety Institute have begun advocating for safety cases before frontier models ship — applying a tool engineers have trusted for decades to software that nobody fully understands yet.