Two recent and contradictory trends in public policies are the increased reliance on the 'dangerous standard' as the primary or sole justification for civil commitment in the mental health system and the decreased reliance on predictions of violence in determining release from prison in the criminal justice system. The most striking conclusion to emerge from the eight major research efforts attempting to validate predictions of violence is the great degree to which violence is overpredicted. Violence is vastly overpredicted whether simple behavioral indicators or sophisticated multivariate analyses are employed and whether psychological tests or thorough psychiatric examinations are performed. Factors accounting for this overprediction include a lack of corrective feedback to the predictor, differential consequences to the predictor of overpredicting and underpredicting violence, illusory correlations between predictor variables and violent behavior, and others. The paper recommends that research on violence prediction employ multiple definitions of violence, multiple time periods for followup validation, and multiple methods of verifying the occurrence of violent behavior and should stress actuarial rather than clinical methods. Appropriate research strategies for each recommendation are detailed. Footnotes and 84 references are supplied.