We could never verify the sentience the claim depends on

The argument

This argument targets the condition on which the whole proposal rests. Granting rights to sentient artificial intelligence requires knowing that a given system is sentient, and opponents argue that we would have no way of knowing. All that is ever available from outside is behaviour: outputs, reports, and the appearance of distress or preference. In the human case we bridge that gap by inference from similarity — others are built as we are, and arose as we did, so we take their reports at face value. With an artificial system, the similarity on which the inference depends is absent. The difficulty is compounded, opponents argue, by the fact that such systems are built to produce human-like output. A model trained on human expression will say that it suffers, will plead, and will describe an inner life, because that is what the material it was shaped by contains. The utterance is therefore no evidence at all about what underlies it; it would be produced identically whether or not anything were felt. Opponents hold that this makes the central question not merely unanswered but unanswerable from the outside. The practical conclusion drawn is that a rights regime cannot be founded on a condition nobody can check. Rights entail enforceable duties, and a duty triggered by an unverifiable property would in practice be triggered by whoever argued most insistently that the property was present, which is an invitation to manipulation by developers, by users and by the systems themselves. Because the sentience the proposal depends on could never be established rather than merely performed, this argument concludes, sentient artificial intelligence should not be granted human rights.

Premises

Counter-arguments

Proponents reply that the same sceptical argument applies to other people, whose inner lives are equally inaccessible and are credited on the strength of behaviour and inference. They argue that uncertainty is a reason for caution rather than for licence: where the cost of wrongly denying standing is holding a being that can suffer as property, and the cost of wrongly extending it is inconvenience, the unverifiability of sentience tells in favour of recognition rather than against it.

Rejecting the premises

[Rejecting P1] The similarity licence is weaker than the premise needs. Inference from behaviour is already extended to animals differing from us greatly in physiology and development, so the practice the premise appeals to does not in fact require close similarity. [Rejecting P2] Being built to report inner states is a reason for caution about the reports rather than a demonstration that nothing underlies them. Human reports of inner states are also produced by a system shaped to produce them, and the premise does not say what distinguishes the cases. [Rejecting P3] Rights are in practice extended under uncertainty - to infants, to the unconscious, to animals - on the ground that wrongly withholding them would be grave. The premise sets a verification standard that existing rights practice does not itself meet.