Aligning AI with helpfulness destroys its human simulation ability
a study of qwen, llama, and olmo models found that post-training makes them worse at predicting human behavior, undermining their use as digital twins for research.
researchers using psych-201, a dataset of 26 million responses, found that the post-training process makes models like qwen3 and llama3 worse at simulating human behavior. base models consistently outperformed instruction-tuned and reasoning variants in predicting survey responses, with the gap widening in newer generations. this indicates that rlhf and alignment techniques systematically overwrite the human quirks and biases necessary for accurate behavioral modeling. the study also found that providing demographic context to models does not mitigate this accuracy loss.