We study whether narratives, when the underlying statistical problem is fixed, distort belief updating. In a between-subjects design, participants solve an isomorphic two-state Bayesian inference problem either in an abstract two-urn task or in a crime story in which they infer which of two suspects is guilty. Both environments feature diagnostic and neutral (nondiagnostic) signals, with likelihoods common knowledge and symmetric. Across ten incentivized sequential rounds, belief reports depart from Bayes in both treatments, consistent with noisy inference, but departures are systematically larger under narratives and statistically significant on standard measures of updating error. The strongest divergence concerns neutral information: in the narrative treatment, neutral signals elicit non-normative updating and a systematic pull toward the focal midpoint 0.5 (“reset to uncertainty”), rather than the Bayes-prescribed inaction. These results contribute to experimental work on belief updating and the economics of narratives by identifying a targeted, welfare-relevant distortion in the treatment of neutral signals.