Humans combine co-emitted social signals to predict other’s immediate intentions and prepare an adapted response. However, little is known about whether attending to only one of co-emitted social signals impacts on its combination with other signals. Here, using electroencephalography, we address selective attention effects on early combination of social signals. We manipulated three visual cues: gaze direction, emotional expression, and pointing gesture, while participants performed either emotion or gaze direction judgments. Results showed that a temporal marker of social cues integration emerges 170 ms after the stimulus onset, even if the integration of the three visual cues was not required to perform the task, as only one feature at a time was task-relevant. Yet, in addition to common temporal regions, the relative contribution of specific neural sources of this integration changed as a function of the attended feature: integration during emotion judgments was mainly implemented in classical limbic areas but in the dorsal pathway during gaze direction judgments. Together, these findings demonstrate that co-emitted social cues are integrated as long as they are relevant to the observer, even when they are irrelevant to the ongoing task.