Evolutionary theory predicts that observable traits should evolve to reliably indicate unobservable behavioral tendencies in coordination games but not social dilemmas. We conducted a two-part study to test this idea. First, we recorded 60-s videos of participants, and then these participants played a stag hunt game or a prisoner’s dilemma. Subsequently, raters viewed these videos, with the sound either off or on, and they guessed player choices. Raters showed a significant tendency to guess that attractive players chose stag. In contrast to the prediction, rater accuracy was at chance regardless of whether the sound of the video was off or on. For prisoner’s dilemma players, raters showed a significant tendency to guess that women cooperated at a higher rate than men. Again in contrast to the prediction, accuracy was significantly above chance in this case. To calibrate the importance of this accuracy rate, we developed two models that suggest the accuracy we observed in the prisoner’s dilemma case is probably not high enough to support the evolution of cooperation. Altogether, our results show that raters tried to achieve a meaningful degree of accuracy about players by using the limited information available in the videos, but they could not do so.