The four models make different trade-offs. Some assume the brain samples only a few possibilities before answering; others introduce biases toward familiar prototypes or add cautious priors that regularize extreme beliefs. By testing these ideas across many published experiments and diverse causal networks, the researchers show which assumptions help reproduce empirical quirks, and where each account falls short. The findings favor a pluralistic perspective in which multiple approximations can be adaptive depending on uncertainty and task structure.
For readers interested in human potential and inclusion, this work matters because it reframes apparent mistakes as rational responses to real constraints. That perspective can change how we teach causal reasoning, design decision aids, and build tools that support people facing complex, uncertain problems. Follow the link to explore how models differ when causes inhibit rather than excite outcomes, and what that tells us about guiding better reasoning under pressure.
Abstract
Human causal judgments frequently deviate from normative Bayesian expectations, particularly with respect to conditional independence and explaining away. Rather than interpreting these deviations as reasoning errors, recent computational accounts suggest they may emerge from principled approximations to ideal Bayesian inference. We evaluate four leading frameworks: Bayesian sampler (BS), mutation sampler (MS), Bayesian mutation sampler (BMS), and Bayesian uncertainty model (BUM), which each formalize different cognitive constraints, including limited sampling, prototype anchoring, prior regularization, and uncertainty over causal structure. These models were systematically compared across 33 reported experimental conditions (N = 1154) spanning diverse causal structures, including common cause, common effect, and chain networks with generative and inhibitory relations. All four models had at least some success reproducing the positive and negative Markov violations observed in common cause and chain structures. In common effect structures with inhibitory or mixed links, only MS and BMS captured the observed patterns more consistently, although BMS’s additional parameter offered limited improvement over MS. BS and BUM showed poorer fits in those conditions, though they may still offer plausible accounts under different assumptions. We discuss these findings in light of prior work on causal reasoning, emphasizing that deviations from normative inference may reflect adaptive strategies shaped by structural uncertainty and cognitive constraints. This supports a pluralistic and resource-rational view of causal reasoning and underscores the need for targeted experiments probing inhibitory causal relations and model uncertainty.
