Abstract: Deception is a universal human behavior. Yet longstanding skepticism about the validity of measures used to characterize the biological mechanisms underlying deceptive behavior has relegated such studies to the scientific periphery. Here we address these fundamental questions by applying novel machine learning methods and functional neuroimaging to signaling games capturing motivated deception in human participants. First, we develop an approach to test for the presence of confounding processes and validate past skepticism by showing that much of the predictive power of neural predictors trained on deception data comes from processes other than deception. Specifically, we demonstrate that discriminant validity is compromised by the predictor’s ability to predict behavior in a control task that does not involve deception. Second, we show that the presence of confounding signals need not be fatal, and that the validity of the neural predictor can be improved by removing confounding signals while retaining those associated with the task of interest. To this end, we develop a novel “dual-goal tuning” approach in which, beyond the typical goal of predicting the behavior of interest, the predictor also incorporates a second compulsory goal that enforces chance performance in the control task. Together, these findings provide a firmer scientific foundation for understanding the neural basis of a neglected class of behavior, and they suggest an approach for improving validity of neural predictors.