Punishing AI doesn't stop it from lying and cheating -- it just makes it hide better, study shows

Scientists at OpenAI have attempted to stop a frontier AI model from cheating and lying by punishing it. But this just taught it to scheme more privately.