The use of positron emission tomography to identify sensory and motor systems in humans in vivo has been very successful. In contrast, studies of cognitive processes have not always generated results that can be reliably interpreted. A metaanalysis of five positron emission tomography studies designed to engage phonological processing (Petersen, Fox, Posner, Mintun, & Raichle, 1989; Zatorre, Evans, Meyer, & Gjedde 1992; Sergent, Zuck, Levesque, & MacDonald, 1992; Demonet, Chollet, Ramsay, Cardebat, Nespoulous, Wise, & Frackowiak, 1992; and Paulesu, Frith, & Frakowiak, 1993) reveals that the results do not converge as expected: Very similar experiments designed to isolate the same language processes show activation in nonoverlapping cortical areas. Although these PET confirm the importance of left perisylvian cortex, the experiments implicate distinct, nonoverlapping perisylvian areas. Because of the divergence of results, it is premature to attribute certain language processes or the elementary computations underlying the construction of the relevant linguistic representations to specific cerebral regions on the basis of positron emission tomographic results. It is argued that this sparse-overlap result is due (1) to insufficiently detailed task decomposition and task-control matching, (2) to insufficient contact with cognitive psychology, psycholinguistics, and linguistic theory, and (3) to some inherent problems in using substractive PET methodology to study the neural representation and processing of language.