Research into the evaluation of transitive inferences from inclusion relations has shown, though there are important intersubject differences, that participants tend to consider the inclusion relation to be symmetrical and that they fail to assume transitivity: the level of correct responses falls for true inferences and increases for false inferences, the greater the number of inferential steps involved (Truth x Distance interaction). This article puts forth a cognitive load hypothesis to account both for the interaction and for intersubject differences. It was shown that the cognitive load associated with the calculation of the inferences predicted error rates and accounted for the Truth x Distance interaction. Moreover, the performances were correlated with participants' working memory span. Research on deductive reasoning in adults has often led to the conclusion that human reasoning does not conform to the rules of logic and is subject to multiple biases (Evans, 1982, 1989). As pointed out by Johnson-Laird and Byrne (1991), "People are rational in principle, but fallible in practice" (p. 19). Numerous theories of reasoning suggest that the limitation of the processing capacities of human participants represents a major source of error (e.g., Braine, 1990; Johnson-Laird & Byrne, 1991; Rips, 1983). The purpose of this article is to determine the extent to which the errors observed in a transitive inference evaluation task known for its difficulty (the set-inclusion task) can be explained in terms of a problem of cognitive load and whether interparticipant differences are due, at least in part, to differences in processing capacity. The most frequently studied field has been that of the comprehension of the conditional (if p then q), in particular using Wason's (1966, 1968) selection task. Although some models point to the implementation of rules of inference