Game theoretical analysis is used to extend the predictions of classical signal detection theory to situations in which the detection task is performed by two observers. The implications of the suggested generalization are demonstrated and examined in the context of a two-person safety game. In a controlled experiment, 28 dyads performed a letter classification task. In accordance with the predictions, naive subjects were less likely to detect critical signals when the responsibility of detection was shared, and the magnitude of the criteria shift depended on the subjects' sensitivity (d'). It is shown that a simple reinforcement-based learning model can approximate the observed results. This model is equally applicable to both the classical one-person and the two-person signal detections tasks.