A large-scale study covering 208,000 participants and 26 million responses shows that the very training that turns language models into helpful chatbots weakens their ability to replicate human behavior. The effect gets worse with each new model generation. Even the popular persona trick, feeding mo
