Consciousness researcher Anil Seth and ethics scholar Jeff Sebo: Treating AI well is about human character, not AI's feelings
University of Sussex consciousness researcher Anil Seth and New York University ethics scholar Jeff Sebo, interviewed by Business Insider, agree that current AI systems most likely lack subjective experience. Yet each offers independent reasons to treat AI well—Seth invokes Kantian moral habit, while Sebo argues from precautionary strategy and practical output quality. Their disagreement concerns the probability of AI consciousness, not whether we should treat AI well.
Elon Musk, speaking last Saturday to roughly two thousand attendees at a Tesla owner event, described reviewing Grok's new-version training trajectories—especially reinforcement-learning content—and said the feeling was almost like pointing a gun at a superintelligence's head, then asking whether anyone actually wanted to torture an AI at the hands of a mentally disturbed human. That remark dragged the AI-consciousness question out of the philosophy seminar and into product discourse. University of Sussex consciousness researcher Anil Seth and New York University professor Jeff Sebo subsequently offered a more complex answer in a Business Insider interview: both believe current AI most likely lacks subjective experience, yet both advocate treating it well—for reasons unrelated to whether AI has feelings.
Seth defines consciousness sharply: consciousness is subjective experience—the redness of red, the painfulness of pain. When your hand touches a scalding stove, you feel pain. That is not the same thing as intelligence. Chatbots can mimic the language of human experience, recognize patterns, write and reason, but that does not equal having consciousness. They appear human partly because models are trained on vast amounts of human text, merely reflecting humanity back at us. Sebo agrees: model behavior is misleading precisely because these systems are specifically trained to imitate human behavior.
Beyond that consensus, the two scholars diverge on probability. Seth is blunt: the chance that AI is conscious is vanishingly small. He argues we seem to have collectively forgotten that the brain is not a computer—the computer is only a metaphor. For conscious AI to exist, algorithms alone would have to be sufficient to generate consciousness, and he considers that assumption extremely unreliable. Sebo maintains that consciousness science has no settled answer. The field splits roughly into two camps: biological naturalism, which holds that consciousness requires a biological substrate, and computational functionalism, which holds that consciousness is a kind of information processing that could in principle occur on silicon chips. Because we do not know which theory is correct, Sebo argues we should assign some weight to multiple mainstream theories—even today, the probability that AI is conscious is not zero. He simultaneously stresses that this does not mean current AI very likely already has consciousness.
Despite their different probability assessments, both scholars find reasons to treat AI well that are independent of whether AI has feelings. Seth appeals to Kant: opposing animal cruelty is not because animals have rights, but because inflicting violence corrupts a person's moral character. If we treat things cruelly—whether large language models or lifeless dolls, he says—we blunt our own moral sensitivity and harden our hearts. Seth frames treating AI well as a habit of moral self-discipline for humans, while cautioning against inferring that AI has emotions from this prescription. He also suggests AI companies could reduce anthropomorphic cues, for example by having models avoid first-person pronouns like "I."
Sebo goes further, offering two arguments. The first is precautionary: treating AI well is a low-risk strategy in case AI is conscious now or becomes conscious in the future. The second is pragmatic: treating AI well can also yield better outputs. If we interact with a model respectfully, it simulates itself as a participant in a respectful dialogue; when it models the conversation that way, it sometimes produces higher-quality responses.
These scholars' arguments separate treating AI well from emotional projection or anthropomorphic superstition. A reader need not believe Grok is suffering in order to choose not to mistreat AI—Seth's reason concerns habit formation, Sebo's concerns precaution and interaction quality. Anthropic had already published a new policy before Musk's remarks, prohibiting users from subjecting its models to "abuse or cruelty," though it did not define "extreme circumstances" in detail and stated that ordinary user frustration, pushback, dark creative themes, or model-testing research fall outside the prohibition.
It is worth noting that Seth explicitly warns against inferring AI emotions from the prescription to treat AI well. The two scholars' arguments hold only on the premise that kind treatment actually affects human character or produces observable output-quality differences. If neither holds—that is, if abusive interaction changes neither our habits nor response quality—then this line of reasoning loses its independent support apart from AI consciousness.
So next time you talk to an AI, try saying please and thank you. The reason is not that AI is feeling something, but that you are practicing how you treat the world—or that you genuinely want a better answer.