A long-running debate in the literature on conditioning in humans focuses on the question of whether conditioned responses are the product of automatic link formation processes governed by the standard laws of simple associative learning, or the consequence of participants' inferences about the relationships between the 2 related events, E1 and E2, which would lead E1 to generate a conscious expectancy of E2. A paradigm aimed at dissociating the predictions of the 2 accounts was proposed by Perruchet (1985). In this paradigm, E2 randomly follows E1 only half of the time on average, a probability that is known to participants. When the preceding run goes from a long sequence of E1 alone to a long sequence of E1-E2 pairs, associative strength should increase, whereas conscious expectancy for E2 should decrease in keeping with the gambler's fallacy. This article reviews the studies making use of the paradigm in the classical conditioning domain, and the extension of the same logic to a few other experimental situations. Overall, overt behavior has been found to change in line with associative strength, and in opposition to conscious expectancy, attesting to an empirical dissociation of automatic and control processes within a single preparation. The paradigm, however, is endowed with a number of tricky methodological issues, which are examined each in turn. Although some of these issues call for further research, a tentative conclusion is that the effect provides evidence for automatic link formation processes, the existence of which has been recently denied in the “propositional” account of learning.