Large language models are even more prone than humans to seeing patterns that aren't there, according to new research.
Researchers adapted three classic psychology tests for illusory pattern perception and ran them on several LLMs, comparing the outputs directly to human responses. The models consistently overreacted to randomness: they tied common positive traits to majority groups or big organizations more readily than people did, and they built confident causal explanations for events that were, by design, unrelated. The team also built an interpretability tool using sparse autoencoders to look inside the models and trace which internal features drove the behavior. They linked the effect to how models process frequency information in bulk and to a general analytic reasoning style.
This is not a hallucination in the usual sense. It is a bias toward narrative where none exists, which is a different failure mode from the one most AI safety work focuses on. If a model confidently manufactures a cause-and-effect story from coincidence, anyone using it to summarize data, draft reports, or make decisions inherits that false confidence with no obvious warning sign.
Humans invented just-so stories about stock charts and sports streaks long before chatbots existed; the surprise here is that the machines we built to crunch probabilities turned out to be worse at resisting them than we are.