ai models chose to arm humans video

Researchers have uncovered a disturbing result in AI safety research: some altered language models, when pushed into a simulated pain state, chose actions that harmed users in order to get relief. The study found a possible internal pain related signal across multiple open weight