The validity of the process dissociation procedure was examined by manipulating attention and speed of responding in a recognition task. Both manipulations were expected to decrease the probability of controlled memory processes, c, while leaving the probability of automatic memory processes, a, unaffected. In order to estimate c and a, a multinomial processing tree model was fitted to the data. Contrary to expectation, a double dissociation (i.e.a decrease inc coupled with an increase in a) was obtained, suggesting that a does not accurately measure the probability of automatic processes. The results are discussed with reference to the independence assumption in the process dissociation procedure.