This paper reports from a series of large-scale survey experiments aimed at identifying whether people have preferences over the social identity of information sources. I examine both naturally occurring identities (caste and religion in India), and experimentally assigned identities (in a EU/US sample). The results show that when the quality of information is known, the identity of the messenger does not influence learning. Further, participants react strongly to signals of information quality in all settings and descriptive results suggest that people may rely on pre-existing beliefs about the abilities of different identity groups in the absence of quality signals. Finally, I show that people prefer to learn from non-social sources (a computer algorithm) than from other people. Taken together, the results suggest that experts and policymakers should prioritise emphasising the quality of information.