Recent work found that people can predict their scores on the Implicit Association Test with remarkable accuracy, challenging the traditional notion that implicit attitudes are inaccessible to introspection and suggesting that people might be aware of these attitudes. Yet, there are major open questions about the mechanism and scope of these predictions, making their implications unclear. Notably, people may be inferring their attitudes from externally observable cues (e.g., in the simplest case, their demographic information) rather than accessing them directly. This problem is exacerbated by the fact that, in past work, predictions have been obtained only for a small set of targets, attitudes toward which are demonstrably easy to infer. Here, in an adversarial collaboration with 5 pre-registered studies (N = 4,541), we interrogate implicit attitude awareness using more stringent tests. We demonstrate that people can predict their implicit attitudes (a) across a broad range of targets (many of which are plausibly difficult to infer without introspection), (b) far more accurately than third-party observers can on the basis of demographic information, and (c) with similar accuracy across two different widely used implicit measures. On the other hand, predictive accuracy (a) was lower than previously reported, (b) varied widely across individuals and attitude targets, and (c) was indeed partially explained by inference over observable cues such as demographic variables or explicit attitudes. Taken together, these findings suggest that successful predictions of one’s own implicit attitudes likely emerge from multiple mechanisms, including inference over observable cues as well as genuine introspective access.