Enabling people to reduce their consumption of unhealthy appetitive products can improve their health. Over the last decades, progress has been made by uncovering new ways to change behavior toward appetitive products without feedback incentives (e.g., reward or punishment, as in feedback-driven reinforcement learning), but instead by cueing motor responses (e.g., go vs. no go) toward these products in cognitive training tasks. However, it is unclear how this nonreinforced learning compares to reinforcement learning. Moreover, recent work on reinforcement learning has uncovered a basic learning mechanism, the action-valence asymmetry, which points to the possibility that reward and punishment learning may not always outperform learning without any external reinforcement. Here, we report two well-powered preregistered experiments (experiment 1a: N = 72
experiment 1b: N = 81) that examined when reinforcement learning outperforms nonreinforced learning in modifying people's preferences for food. Our findings show that reinforcement learning notably surpasses nonreinforced learning, but only when active responses (go) are rewarded, and inactions (no-go) are reinforced by avoiding punishments. These results shed light on interventions that combine rewards and punishments to facilitate changes in food preferences.