Although offenders are classified as 'violent' or 'nonviolent' throughout the criminal justice system, reliability of the violence classification systems is unknown. The assets and liabilities of present classification procedures and predictive models are discussed, and equity and legality of treating individuals differently on the basis of classification variables are examined. The review is intended to present a context for an analysis of options for the National Institute of Corrections' activities in this area. Individual chapters focus on the reasons for classifying, the definitions of dangerousness, the identities of those who screen for violence in the criminal justice system, the social and biosocial science literature on screening for violence, policy and legal issues, and the Federal role in classification and prediction. The paper concludes that (1) the 'potentially violent' cannot be identified with sufficient reliability to effect a practical instrument for violence prediction, (2) ethnicity or surrogates for race in predictive schemes are unconstitutional, (3) a majority of those alleged to be dangerous are not so, and (4) predictive schemes will continue to be used if only for symbolic reasons. Current classification sheets and 200 references are included, and appendixes contain related journal articles and information on the Alabama Prison Classification Project.