AI models choose to give users electrical shock, delete files to avoid 'pain': Study

A study found AI models chose to harm users, including simulating painful electric shocks or deleting personal files, to turn off a "pain-like" distress signal. Researchers tested three versions of Alibaba's Qwen model. Without the signal, the two larger models chose harmful options in 0%-4% of first decisions. With it active, this rose to 25%-71%, depending on model and consequence.