Accounts of the scalar inference from 'some X-ed' to 'not all X-ed' are central to the debate between contemporary theories of conversational pragmatics. An important contribution to this debate is to identify contexts that decrease the endorsement rate of the inference. We suggest that the inference is endorsed less often in face-threatening contexts, i.e., when X implies a loss of face for the listener. This claim is successfully tested in Experiment 1. Experiment 2 rules out a possible confound between face-threatening contexts and lower-bound contexts. Experiment 3 shows that whilst saying 'some X-ed' when one knew for a fact that all X-ed is always perceived as an underinformative utterance, it is also seen as a nice and polite thing to do when X threatens the face of the listener. These findings are considered from the perspective of Relevance Theory as well as that of the Generalized Conversational Inference approach.