Abstract
The amount of fear evoked by potential threats is oftentimes proportional to the overlap in shared features with known threats. An adaptive learning system should therefore extract relevant features from threat stimuli to successfully detect other novel threats in the environment. But what if the most relevant feature of a threat stimulus is emotionally positive? Here, we used Pavlovian fear conditioning to ask whether people extract positive emotional features of a fear conditioned stimulus (CS) to selectively generalize to other stimuli that contain positive features. In a between subjects design, we first paired a picture of a face expressing either a slight amount of happiness or fear with an electrical shock to the wrist. We then tested fear generalization to modified face stimuli of the same identity expressing more or less happiness or fear. Both groups exhibited biased physiological arousal (a peak shift) to a face stimulus with the most exaggerated emotional expression, regardless of valence. Fear generalization diminished to unreinforced happy faces over the course of testing, whereas arousal was maintained to unreinforced fearful faces throughout testing. Finally, subjects fear conditioned to a slightly happy face were accurate at retrospectively identifying the correct CS, whereas subjects fear conditioned to a fearful face retrospectively misidentified a more fearful face as the threat CS. These findings suggest that positive emotional features extracted from a known threat can guide biased fear generalization, but that generalization is maintained by a dimension of increasing fear which also produces retrospective biases in threat intensity estimation.