How should an algorithm deal with the danger of making a wrong decision? The current work addresses this question in an experiment. In it, subjects set a parameter which controls a data-driven algorithm. This parameter, a threshold, determines whether a person’s preference for risk is high enough for a financial risk to be taken. A lower threshold means that the algorithm implements risky choices more often and that the rate of false positives is higher. Conversely, a higher threshold implies less risky choices but also more false negatives. The findings are as follows: i) When setting the threshold for themselves, subjects leave the algorithm little influence in the eventual decisions by either setting the threshold very high or very low. When setting it for others, they leave the algorithm more influence. On average, however, chosen thresholds are the same when subjects set it for decisions which affect themselves as when they set it for decisions which the algorithm makes for others. ii) Subjects respond to conflicts of interests: When setting the threshold for decisions which affect others, subjects respond to a bonus for inducing risky decisions. They do so by setting the threshold lower than when the bonus is absent. iii) Threshold choices in all treatments are unaffected by whether the choice procedure is framed as being based on a “computer algorithm” or a human-developed “decision rule”.