Functional magnetic resonance imaging (fMRI) is a potential paradigm shift in psychiatric neuroimaging. The technique provides individual, rather than group-averaged, functional neuroimaging data, but subtle methodological confounds represent unique challenges for psychiatric research. As an exemplar of the unique potential and problems of fMRI, we present a study of 10 inpatients with schizophrenia and 10 controls performing a novel "n back" working memory (WM) task. We emphasize two key design steps: (1) the use of an internal activation standard (i.e., a physiological control region) to address activation validity, and (2) the assessment of signal stability to control for "activation" artifacts arising from unequal signal variance across groups. In the initial analysis, all but one of the patients failed to activate dorsolateral prefrontal cortex (DLPFC) during the working memory task. However, some patients (and one control) also tended to show sparse control region activation in spite of normal motor performance, a result that raises doubts about the validity of the initial analysis and concerns about unequal subject motion. Subjects were then matched for signal variance (voxel stability), producing a subset of six patients and six controls. In this comparison, the internal activation standard (i.e., motor activation) was similar in both groups, and five of six patients, including two whom were neuroleptic-naive, failed to activate DLPFC. In addition, a tendency for overactivation of parietal cortex was seen. These results illustrate some of the promise and pitfalls of fMRI. Although fMRI generates individual brain maps, a specialized survey of the data is necessary to avoid spurious or unreliable findings, related to artifacts such as motion, which are likely to be frequent in psychiatric patients.