Our brains predict future sensory input based on their current beliefs about the world around us, but interpreting prediction errors can be challenging in a volatile environment because they can be caused by stochastic sensory noise or by outdated predictions. Noisy signals should be integrated with prior beliefs to improve precision, but the two should be segregated when environmental changes render prior beliefs irrelevant. Bayesian inference provides a statistically optimal solution to deal with such problems of causal uncertainty [1]. However, the method quickly becomes memory intensive and computationally intractable when applied consecutively [2]. Here, we systematically evaluate the predictive performance of consecutive Bayesian causal inference for human perceptual decisions in a spatial prediction task based on noisy audiovisual sequences with occasional changepoints [3]. We elucidate the simplifying assumptions of a previously proposed reduced Bayesian observer model [4] and we compare it to an extensive set of models based on alternative simplification strategies. Model-free analyses revealed the hallmarks of Bayesian causal inference: participants integrated sensory evidence with prior beliefs in a reliability-weighted fashion to improve accuracy when prediction errors were small, and prior weights gradually decreased as prediction errors grew larger, signalling probable irrelevance of the prior due to a changepoint. Model comparison results further supported the hypothesis that participants computed probability-weighted averages over both causal options (noise or changepoint) after every stimulus, rather than deterministic selection of either cause or keeping both options in memory. Participants thus iteratively summarized their beliefs while accounting for causal uncertainty akin to the reduced Bayesian observer model with minimal memory capacity. However, we also found that participants’ reliance on prior beliefs was systematically smaller than predicted by the model, and this was best described by individually fitting lower-than-optimal parameters for the a-priori probability of prior relevance. We conclude that perceptual belief updating in volatile environments with stochastic noise is best described by a simplified model of consecutive Bayesian causal inference. Observers utilize priors flexibly to the extent that they are deemed relevant, though also conservatively by trusting them less than an ideal Bayesian observer.