Abstract
The surge of online scams is taking a considerable financial and emotional toll. This is partially because humans are poor at detecting lies. In a series of three online experiments (N(exp1) = 102, N(exp2) = 108, N(exp3) = 100) where participants are given the opportunity to lie as well as to assess the potential lies of others, we show that poor lie detection is related to the suboptimal computations people engage in when assessing lies. Participants used their own lying behaviour to predict whether other people lied, despite this cue being uninformative, while under-using more predictive statistical cues. This was observed by comparing the weights participants assigned to different cues, to those of a model trained on the ground truth. Moreover, across individuals, reliance on statistical cues was associated with better discernment, while reliance on one's own behaviour was not. These findings suggest scam detection may be improved by using tools that augment relevant statistical cues.