TRUTH BIAS IN HUMAN REINFORCEMENT LEARNING

Truth bias, the inclination to accept information as true even when cues of deception are present, is a well-documented phenomenon in social and cognitive psychology. However, much of this evidence derives from belief reports concerning abstract beliefs shaped by complex ideological priors and episodic memory, while its influence on instrumental reinforcement learning (RL) remains underexplored. We investigated this using a disinformation two-armed bandit task where 201 participants repeatedly chose between two options to maximize reward outcomes. They received outcome-feedback about their choices from sources with different propensities for lying (0%, 33%, 67%, or 100%; explicitly instructed to participants). When a source lied, the feedback was the opposite of the actual reward-outcome (Fig. 1a-c). This ensured that every truthful source (e.g., 100%) had an equally informative deceitful source (e.g., 0%). We applied computational RL modeling to capture how participants integrated this feedback. The model estimated a "credit assignment" parameter for each source-feedback, reflecting the magnitude of belief updates based on such feedback. Our computational modeling revealed a clear truth bias: participants learned significantly more from truthful feedback compared to deceitful feedback (Fig. 1d). Furthermore, since the feedback source's honesty was forewarned before each choice (Fig. 1c), we could examine how expected truthfulness influenced decisions. We found that participants were both less accurate (Fig. 1e) and slower (Fig. 1f) in their choices when they expected to receive deceitful (rather than truthful) feedback. Our findings demonstrate a deeply ingrained truth bias that influences both fundamental learning mechanisms and decision-making behavior within reinforcement learning.