Because the world is dynamic and only imperfectly observable, many of the decisions we make are necessarily uncertain. How do we navigate such uncertainty? From the perspective of cognitive neuroscience, the classic answer would be that we evaluate the benefits of each potential choice and then lean towards the one promising the greatest reward, modulo some exploratory noise. Conversely, an ethologist would argue that we would stay with previously rewarding choices until the payout drops below a certain threshold, at which point we start exploring other options. While both hypotheses wield considerable influence within their respective fields, it remains uncertain which one best describes human decision-making. Here, we asked whether human decision-making was better described as a compare-to-threshold process or as a value-comparison process in a classic testbed of decision-making under uncertainty from the reinforcement learning (RL) literature: a restless k-armed bandit task. We found that human behavior more closely resembled compare-to-threshold computations than value-comparison computations. This insight is difficult to reconcile with traditional reinforcement learning models, which center value-comparison computations. Therefore, we next developed a novel compare-to-threshold (“foraging”) model and asked whether it better explained participants’ behavior in multiple RL tasks. The foraging model was a better fit for participants’ behavior and better predicted the participants' tendency to repeat choices on both individual and group level. The foraging model was also able to predict the existence of held-out participants with very prolonged repetitive choice runs, a pattern that was almost impossible in the RL model. These findings indicate that humans use compare-to-threshold computations---even in tasks that were designed as testbeds for RL algorithms. They also offer a novel cognitive account of decision-making under uncertainty with widespread implications for the cognitive and brain sciences.