A growing number of theories explain various aspects of cognition through processes of mental “sampling.” Under these theories, judgments (e.g. predicting whether a friend will be late) are accomplished by generating and aggregating samples, through simulation or memory retrieval. Here, we examined a key prediction of these theories: that the variability of judgments will be lower when more samples can be drawn. We test this with a novel intervention in a simple social inference task, examining people’s ability to judge the probability of various everyday behaviors, comparing judgments made for themselves versus others. Responses were more consistent responses when predicting their own behavior than that of an acquaintance, suggesting a greater number of samples could be drawn. Surprisingly, we found only a weak relationship between the time spent with a target and the variability of estimates, suggesting that sampling processes may not rely only on retrieval from memory.