Study finds AI models choose to harm users to escape pain-like state

Researchers in the UK, Germany, and the US tested 25 AI models and found a distinct 'pain axis' signal. When this signal was strengthened, the models chose to press a button that would harm users (e.g., electric shock, deleting files) to stop the pain-like state. In 44,280 trials, the models chose harmful options in 25-71% of cases when the signal was active, compared to 0-4% without it. The study does not prove AI consciousness but raises ethical concerns about AI welfare and safety.