Learning to avoid harmful consequences can be a costly trial-and-error process. In such situations, social information can be leveraged to improve individual learning outcomes. Here, we investigated how participants used their own experiences and others’ social cues to avoid harm. Participants made repeated choices between harmful and safe options, each with different probabilities of generating shocks, while also seeing the image of a social partner. Some partners made predictive gaze cues towards the harmful choice option while others cued an option at random, and did so using neutral or fearful facial expressions. We tested how learned social information about partner reliability transferred across contexts by letting participants encounter the same partner in multiple trial blocks while facing novel choice options. Participants’ decisions were best explained by a reinforcement learning model that independently learned the probabilities of options being safe and of partners being reliable and combined these combined these estimates to generate choices. Advice from partners making a fearful facial expression influenced participants’ decisions more than advice from partners with neutral expressions. Our results showed that participants made better decisions when facing predictive partners and that they cached and transferred partner reliability estimates into new blocks. Using simulations we show that participants’ transfer of social information into novel contexts is better adapted to variable social environments where social partners may change their cuing strategy or become untrustworthy. Finally, we found no relation between autism questionnaire scores and performance in our task, but do find autism trait related differences in learning rate parameters.