Recently adopted statistical approaches improve researchers' ability to describe what is, and what is not, possible in the prediction of violence by psychiatric patients. At the base rates of violence routinely encountered in outpatient settings, current assessment methods would require hospital admission of large numbers of patients who are potential offenders in order to prevent the actual offending of a few. Suggestions that substantially greater accuracy is possible for short-term predictions, for particular symptom clusters, and for particular offenses have yet to be tested and confirmed. Further research may improve this state of affairs, for instance, by concentrating on particular patient groups. There are reasons to suspect that any improvement will be limited. Clinical practice, however, is likely to continue to require the assessment of a patient's potential for acting violently. Future research should aid such assessments by clarifying the mechanisms by which risk factors correlate with violence and by establishing the clinical usefulness of actuarial scales.