Jonathan Birch

Can there ever be a reliable test for AI consciousness?

Behavioural indicators for consciousness in AI face a gaming problem: an AI system may mimic any list of behavioural indicators without possessing the underlying natural kind they indicate in living animals. Meanwhile, theory-driven architectural indicators face a Janus problem: the apparent ease of recreating the relevant features in AI might be taken as a sign not of the imminent arrival of conscious AI but of the inadequacy of current theories. In the case of other animals (especially mammals and birds), this reaction seems much less reasonable, because a rich background of ungamed behavioural and neural evidence has brought us to a position of equipoise or better, and against that background a theory-driven indicator can shift the dial yet further. But that rich background will most likely never exist for AI systems. Is there a way out? The most obvious way out is to leverage our more favourable situation regarding other animals to develop more secure theories of consciousness than we currently have.