Barnabás Molnár^1,2 (), József Fiser^1,2, Ádám Koblinger^1,2; ¹Central European University, ²Center for Cognitive Computation

The conjunction fallacy is a well-known breakdown of Bayesian reasoning, but its cognitive roots remain unclear. Most studies rely on ambiguous, story-based vignettes that leave room for interpretation. Less is known about its emergence in simpler, perception-based tasks with tightly controlled visual stimuli. We aim to explore this gap in two controlled experiments. In Experiment 1, we tested how people learn probabilistic rules involving abstract visual features in a continuous stream. Participants (N = 41) judged the probability of individual and combined features, as in traditional conjunction-error tasks. Individuals could identify typical features for trial types with specific traits, suggesting that representativeness was learned. Conjunction errors were rare; moreover, varying the representativeness of a feature did not significantly affect the fallacy rate, even though representativeness has been argued to be the leading cause of the fallacy. In Experiment 2, we used a more complex stimulus space defined by color, shape, and pattern. Participants (N=56) estimated single and combined feature probabilities using ranking or slider methods. Error rates similar to those observed in the original tasks emerged only under high comparison load, when multiple statements were ranked simultaneously or when participants had to maintain earlier estimates to stay consistent. However, when participants assessed a single conjunction and its constituent, fallacy rates dropped. Task order also had a significant effect: starting with the probability assessment lowered fallacy rates in the sorting task. Together, these experiments offer new insights into the conjunction fallacy in perception-based learning tasks. We found that the conjunction errors in perceptual tasks arise less from representativeness per se and more from constraints imposed by task structure. These findings support the theory of bounded rationality and show that fallacy rates can be reduced by presenting information in a visual sequence, thus reducing cognitive load.