Implicit measures were introduced to explain phenomena that are characterized by a gap between self-reported attitudes and behavior. Recent meta-analyses revealed, however, that implicit measures have only limited predictive validity that goes beyond self-reports. We identify possible reasons for this failure: (a) A lack of validity that is due to the influence of extraneous processes, (b) a focus on evaluation instead of motivation, (c) a focus on associations instead of propositional beliefs, and (d) a focus on global instead of context-dependent attitudes. Recent developments in the field of implicit measures addressed these problems: (a) Statistical process models increase the internal validity of implicit measures, (b) implicit measures of wanting have the potential to predict behavior better than implicit measures of liking, (c) new paradigms provide measures of automatically activated attitudes for propositions that have an unambiguous interpretation, and (d) assessment of context-dependent beliefs is better suited to predict specific behaviors. Incorporating these developments into research on implicit bias will help to realize the initial expectations of describing, explaining, and predicting behavior in many situations.