LLMs fail to simulate human belief formation in social science experiments

The study's core finding isn't just that models fail at belief simulation, but that they fail asymmetrically: they can reproduce final distributions when anchored to human starting positions, yet cannot generate plausible initial stances or trace realistic reasoning paths. This suggests models memorize distributional targets without learning the underlying update mechanisms.

This connects directly to the salience bias work from late July, which showed models systematically discard implicit reasoning in favor of salient surface features. Here we see a related failure mode: models appear to learn what humans conclude, not how they reason toward conclusions. The consciousness-steering paper from the same period also hints at this gap, revealing that safety fine-tuning reshapes model outputs in ways that diverge from human cognition. Together, these three papers suggest current models have learned to approximate human outputs without acquiring human-like reasoning under uncertainty.

If researchers at Prolific or similar platforms release follow-up work testing whether fine-tuning models on explicit belief-update trajectories (not just endpoints) closes this gap, that would indicate the problem is trainable rather than architectural. If the gap persists across model sizes through 2027, it signals a fundamental limitation in how foundation models represent causal reasoning.