A safety case is a structured argument, supported by evidence, that a system is acceptably safe for a specified environment and use. It has long been used in safety-critical industries such as nuclear power, aviation, railways, and defense, where regulators require operators to present one before approving a system for operation.
In AI, safety cases have been proposed as a governance framework in which developers of frontier AI models must show, before training or deploying a model, why it will not cause catastrophic harm. Supporting evidence can include evaluation results showing the absence of dangerous capabilities, control measures that would prevent harm even if the model attempted it, and evidence that the model behaves in line with human intent. Approaches under discussion include developers publishing their own assessments and review by regulators.