It does (probably) work - implicit attitude scores predict a number of interesting behavioral components. We as a field are still trying to figure out exactly how powerful the prediction is.
For example, if you give someone an explicit measure of prejudice against Black persons, and an IAT measuring implicit prejudice against Black persons, something interesting happens.
The explicit measure will predict how biased their conversation with a Black person is, as judged by a blind panel - how hostile they seem, for instance.
The implicit IAT measure, however, will predict how far away they sit, how many nervous stumbles they make in their speech, how much eye contact they make...all the subconscious stuff that the conversational partner can pick up.
Faking them is, as far as I know, next to impossible, barring systematic delays on contrastereotypic associations to exaggerate the bias. I don't immediately see any way to cheat in the reduced-bias direction.