The increasing spread of misinformation ("fake news") represents one of the great challenges digital societies are facing today. One aspect of this phenomenon is that misinformation often comes from unknown sources that conjure the impression of professionalism. Little is known about the individual characteristics underlying susceptibility to such fraudulent sources. We use data from an original survey experiment that tests the role of sources (known vs. unknown), channels (Facebook vs. website) and content (congruent vs. incongruent) on people's beliefs and sharing of information. We investigate which groups are most susceptible, first, to the source effect, that is, the tendency to believe and share information more from a known than from an unknown source; and second, which groups are more likely to believe and share from a source that has previously provided them congruent information. We apply recently developed machine learning methods based on random forests (Wager and Athey 2018) to tease out heterogenous treatment effects over a wide range of covariates. We find strong heterogeneity on some variables (e.g. age, vote choice, education), but less on others (e.g gender).