Delusion-like content emerges from common-cause inference over semantically disparate observations

Delusions are beliefs with extraordinary content, yet computational accounts of delusions have not explained how this content arises. We propose that extraordinary content emerges when an ordinary mechanism, common-cause inference, is applied to semantically disparate observations. We tested this with large language models (LLMs) by systematically manipulating semantic disparity between observations. Single causes inferred by these models from ordinary observations generally resembled true causes, yet became increasingly delusion-like as observations became more disparate. Observation disparity predicted delusionality, and this relationship could be explained geometrically via semantic analogues of common-cause properties. We further tested this idea in a large corpus of real-world delusions. In cases documenting pre-delusional observations, subsequent delusional beliefs were consistent with what LLMs would infer and exhibited corresponding common-cause properties. Together, our findings identify a computational route through which extraordinary delusion-like content can emerge from ordinary causal inference, and demonstrate its relevance to clinical delusions.