The use of risk assessment tools has come to play an increasingly important role in sentencing decisions in many jurisdictions. A key issue in the theoretical discussion of risk assessment concerns the predictive accuracy of such tools. For instance, it has been underlined that most risk assessment instruments have poor to moderate accuracy in most applications. However, the relation between, on the one hand, judgements of the predictive accuracy of a risk assessment tool and, on the other, conclusions concerning the justified use of such an instrument in sentencing practice, is often very unclear. The purpose of this paper is to examine this relation. More precisely, it is argued that the relation between predictive accuracy and the question as to whether a new risk assessment tool should be introduced in sentencing practice is highly complicated. For instance, there may be cases in which a new risk assessment tool is more accurate than those currently in use, but should nevertheless not be introduced; and conversely, where a new tool is less accurate, but where its introduction instead of current tools would be morally preferable.