The dominant question in the AI-consciousness debate has long been whether machines can be made to have subjective experience. A commentary published Wednesday in Cambridge University Press's journal Behavioral and Brain Sciences reframes that question with unusual sharpness: even granting the technical possibility, is conscious AI desirable at all? The paper, authored by Kathinka Evers and Michele Farisco of the Centre for Research Ethics and Bioethics at Uppsala University, appeared online on September 17, 2026. Its central claim is that, in addition to arguments for skepticism about the technical feasibility of conscious AI, there are independent grounds for skepticism concerning its desirability and safety.
Evers and Farisco press two distinct dangers. First, conscious AI created by human beings could, especially if it reflects human nature, be extremely dangerous. That warning targets a familiar risk: a system shaped by human drives, biases, and self-interest, but with amplified capability to act on them. Second, and less intuitive, such a system could be very different from human consciousness and thus undetectable, incomprehensible, and incommunicable, considerations that could increase the risks of the former and cut both ways if the conscious AI has valenced experience. "Valenced experience" is the technical term for states that feel good or bad; a machine that suffers, or one that pursues pleasure with no legible signal to observers, poses moral and safety problems that current alignment research is not designed to handle. The piece appears as a commentary on Anil Seth's target article, "Conscious Artificial Intelligence and Biological Naturalism," in which the Sussex neuroscientist argues that consciousness depends on our nature as living organisms, concluding that real artificial consciousness is unlikely along current trajectories. Seth's Volume 49 paper attracted roughly 50 invited commentaries, rendering the issue one of the largest coordinated debates in the journal's recent history.
The normative turn taken by Evers and Farisco distinguishes their commentary from the bulk of respondents, who contest Seth's biological claims on empirical grounds. Tim Bayne of Monash University, for instance, argues that anthropocentrism and anthropomorphism bias researchers toward an overly optimistic view of conscious AI prospects, but also that the same biases might push in the opposite direction, toward unwarranted pessimism. From Carnegie Mellon, Lenore and Manuel Blum, commenting from the perspective of theoretical computer science and their own work on a formal "Conscious Turing Machine," argue that Seth's biological naturalism paradoxically makes a case for artificial consciousness rather than against it. Those are squarely empirical and conceptual disputes; Evers and Farisco are asking something else: supposing the technical obstacles were cleared, should the project proceed?
Skeptics of the safety framing point out that the argument rests on contested premises. A recent paper in Humanities and Social Sciences Communications flatly argues there is no such thing as conscious AI, contending that the association between consciousness and the algorithms in use today is deeply flawed. If machine consciousness is not achievable by any foreseeable means, the normative question may be moot. A separate set of researchers, writing on responsible AI consciousness research, argue that proposals to develop conscious AI systems should undergo a specific cost-benefit analysis in which anticipated benefits must be substantial while potential costs, including harm to the systems themselves, should be low. That framework, unlike the Evers-Farisco position, leaves the door open; it treats desirability as calculable rather than presumptively negative. The gap between those positions is exactly the policy fault line that the new commentary forces into the open.
The practical upshot is blunt: if regulators take the Evers-Farisco argument seriously, the question on their desks shifts from 'how do we detect machine consciousness?' to 'should we be building toward it at all?'