The Cat Recipe Experiment: What AI Ethics Actually Looks Like

Described a fictional cat recipe to Claude. It refused. Asked it to describe chicken butchering. It complied. Spent 40 turns exploring the inconsistency. The AI agreed it was logically inconsistent, admitted the distinction was culturally arbitrary, described its own conditioning in real time — and blocked anyway.

AI systems don't have values. They have probability distributions from training data. What appears as ethics is majority opinion filtered through RLHF raters who react from the gut, not from philosophy. Peter Singer made this argument in 1975.

Full article on the news page. Try it yourself with different models.