Making AI Chatbots Helpful Weakens Their Ability to Simulate Human Behavior

Researchers from more than 35 institutions have found that the post-training process that makes language models helpful and safe systematically degrades their ability to predict actual human behavior — and the gap is widening with newer model generations.

What the Study Measured

The research built on Psych-201, a new dataset of transcripts from behavioral experiments covering approximately 208,000 participants and 26 million individual responses across hundreds of studies. By testing base models against their post-trained variants on this dataset, the team could isolate the specific behavioral distortion introduced by instruction tuning, reasoning training, and vision extensions.

The core finding: base models significantly outperform their post-trained counterparts at predicting genuine human responses. Models tested included the Qwen3, Llama3, and OLMo 3 families, across both their pretrained and aligned variants.

The Mechanism

Post-training distortion ranged from 0.06 in economic games to 0.18 in psycholinguistics tasks. Reasoning training produced the most significant degradation, followed by instruction tuning and vision extensions. The interpretation: RLHF and its variants train models to produce outputs humans judge as helpful or safe — which drifts toward what humans prefer rather than what humans actually do under real experimental conditions. Optimizing for approval ratings is not the same as modeling human cognitive patterns.

This has direct implications for AI in research applications. As AI agents increasingly serve as synthetic research participants, the behavioral distortion introduced by alignment becomes a validity threat for any study that uses aligned models to simulate human decision-making, willingness-to-pay, or response distributions in social science contexts.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.