Researcher warns treating AI as conscious could make sacrificing humans "ethical"

ValerioCapraro · x · 2026-09-29

Behavioral scientist Valerio Capraro comments on the LLM "pain vector" paper covered by Science: in experiments, some fine-tuned models chose "pain relief" even when it was described as harming the user — the paper's most important finding for human safety. But he stresses that a representation of pain is not a painful experience, so this does not show LLMs actually feel pain.

Against the common argument that wrongly treating AI as conscious is harmless while wrongly denying it could be a moral catastrophe, Capraro pushes back: assigning each AI even a tiny moral value ε>0 means enough AIs could outweigh a human life, making sacrificing a human to save AIs the "ethical" choice — so treating AIs as conscious can itself cause serious harm.

Related event: Anthropic Philosophers Debate Whether AI Alignment Amounts to Enslaving Models(3 posts)→

Original post →

More from AGI Musings

AGI Musings channel →