We provide the first empirical test of a recent, normative account of the conjunction fallacy. According to Bovens and Hartman (2003), an unlikely statement from a partially reliable source is not necessarily more likely than a conjunction statement from another partially reliable source. Hence once information is considered to be coming from potentially not fully reliable sources, the conjunction fallacy is no longer at odds with probability theory. We provide here a simple experimental test of this account, and report comparisons of the Bovens and Hartmann model with Wyer's (1976) model and a simple averaging model. Wyer's model provided the best fit and the averaging model had the highest true positive rate in determining whether individual participants would commit the fallacy or not.