Study: AI Chatbots Lose Humanity as They Become More Useful

A large-scale study conducted by an international consortium, including researchers from Helmholtz Munich, has revealed that the process of transforming raw language models into useful chatbots weakens their ability to mimic human behavior. This trend intensifies with each new generation of models.

Language models are increasingly being used to replace human subjects in various applications, such as predicting reactions to policies, simulating clinical training for psychiatrists, or modeling student learning. However, this study highlights an uncomfortable conclusion: the training steps that make language models practical assistants render them less effective at modeling human behavior.

The study relies on Psych-201, a dataset of transcripts from behavioral experiments, covering approximately 208,000 participants and around 26 million individual responses from hundreds of experiments. This dataset is several times larger than any previous collection of its kind. Each data point captures the complete journey of a participant through an experiment, including detailed metadata such as age, nationality, questionnaire responses, and other characteristics. This vast dataset was assembled through an open research collaboration involving researchers from over 35 institutions.

Base Models Outperform Their Fine-Tuned Counterparts

Researchers compared models from the Qwen3, Llama3, and OLMo 3 families, testing both base models and their various fine-tuned variants. Base models are trained solely to predict the next word in a text. From there, additional training produces fine-tuned versions to follow instructions, engage in step-by-step reasoning, or process images. The metric used was each model's ability to predict the actual responses given by human participants.

Across the three model families, base models demonstrated a better ability to predict human behavior than their fine-tuned assistant versions. This result is consistent across all families and sizes of models. Base models consistently outperform their fine-tuned descendants, and this effect manifests for every current training objective, peaking with reasoning models, followed by instruction tuning and visual extensions. In almost every direct comparison, the base model surpasses its specialized variant.

An obvious counter-explanation could be that assistant models simply respond in a more deterministic manner and fail to capture the natural diversity of human behavior. Researchers tested this hypothesis with a precision analysis on a subset of tasks with discrete response options. Fine-tuned models still performed worse, making it unlikely that increased determinism is the sole explanation.

The Gap Widens with Each Generation

As base models gradually improve from Qwen2 to Qwen2.5 and then to Qwen3, becoming better at predicting human behavior with each generation, the gap with their derived assistant models continues to widen. Ongoing advancements in fine-tuning exacerbate the divergence from human behavior.

The greatest distortion manifests in linguistic tasks and reasoning. Researchers propose a plausible explanation: base models are, at their core, human language models and thus well-calibrated for language processing tasks. Fine-tuning techniques, such as reinforcement learning based on human feedback, steer them away from this initial goal towards more user-friendly or normatively correct responses.

The same occurs with reasoning. Human decisions are shaped by heuristics and systemic biases that base models seem to capture. Reasoning training optimizes for logically correct answers, precisely crushing the human nuances that matter for behavioral simulation.

A Popular Shortcut Doesn't Work

A second finding concerns a widely used technique: providing language models with participant-specific information to place them in a particular role. In the study, this took the form of an interview format where demographic details about each person were added before the experiment. When possible, prompts included age, gender, nationality, education, clinical diagnoses, and questionnaire scores.

Adding real demographic profiles hardly improves predictions of individual behavior. The effect was virtually null. This held true even when the analysis was limited to developmental psychology experiments, where age-related differences should be informative. Previous work had shown that persona prompts could produce response distributions similar to those of humans at the population level. But the new study calls into question their ability to predict individual behavior or merely appear plausible on the surface.

Centaur Shows Targeted Training Can Still Help

The authors consider their results as a variation of a known problem: additional training towards specific goals can degrade the capabilities acquired during pre-training. To test whether this is a strict limit, they examined Centaur—a model specifically fine-tuned on a portion of the behavioral data.

Centaur showed much higher agreement with human behavior even on new tasks that were not part of its training. Thus, additional training can help, but only when it targets behavioral modeling rather than logical correction.

For research practice, the conclusion is clear: practical and readily available assistant models are not automatically the best choice for behavioral simulations. Researchers recommend either raw base models or variants specifically trained for behavioral simulation. The code and data are available on Hugging Face and GitHub.

The fact that chatbot models have their limits as digital test subjects is not new. A recent study on nine open-source language models revealed that optimizing for more human-like outputs comes at the expense of factual accuracy, and a classifier exposed AI responses with an accuracy of 70 to 80%. The persona trick also performed worse than expected.

Another study showed that models can barely pass as weak or strong learners on command, with success rates varying by less than one percentage point. And regarding reasoning, a deep gap persists: an analysis of over 170,000 reasoning traces showed that reasoning models think differently from humans, falling into a kind of sequential autopilot.