Performance-Learning Dissociation in AI-Assisted Teams: Evidence from Inter-brain Synchronization during a Reversal Learning Task

Artificial intelligence (AI) is rapidly transitioning from a passive tool to an active collaborator, raising concerns about its impact on human learning. Prior research has suggested a potential dissociation between task performance and learning, where AI assistance may boost immediate outcomes while hindering deeper cognitive processing. However, the neural mechanisms underlying this dissociation, particularly in team contexts, are unknown. This study investigated how AI, as a team member, affects dyadic performance, learning, and inter-brain synchrony during a probabilistic reversal learning task. We hypothesized that while AI assistance would enhance task accuracy, it would impair learning processes and reduce inter-brain synchronization compared to human-only collaboration.A total of 50 same-sex dyads of university students (N = 100) were randomly assigned to either an AI-assisted group or a no-AI control group. Each dyad completed a collaborative probabilistic reversal learning task, where one member acted as the decision-maker and the other as an advisor. In the AI-assisted group, an AI advisor (accuracy fixed at 65% based on a pilot benchmark) provided an additional suggestion. Behaviorally, we measured decision accuracy, learning rate (trials to stable switching), reversal switching speed (trials to first correct switch after rule reversal), and drift rate (v) using the Drift Diffusion Model. Functional near-infrared spectroscopy (fNIRS) hyperscanning was employed to simultaneously record brain activity from the dorsolateral prefrontal cortex (DLPFC), frontopolar cortex (FPC), and temporo-parietal junction (TPJ) of both participants, allowing for the calculation of regional brain activation, intra-brain functional connectivity (FC) and inter-brain synchronization (IBS).Consistent with our hypotheses, behavioral results showed that the AI‑assisted group demonstrated higher final decision accuracy than the no‑AI group, indicating a performance benefit. However, this benefit did not translate into better learning. The AI‑assisted group exhibited a slower learning rate, required more trials to achieve stable switching after a rule reversal, and showed a significantly lower drift rate, suggesting less efficient evidence accumulation. In contrast, the no‑AI group displayed a negative learning trend across phases in terms of trials to stable switching (meaning the number decreased over time), indicating progressive improvement without AI. Neurally, while the AI‑assisted group showed stronger DLPFC activation, they displayed weaker functional connectivity across frontal and temporal regions, though this difference did not survive multiple comparison correction. Moreover, the AI‑assisted group had significantly reduced inter‑brain synchronization in the DLPFC compared to the no‑AI group, both during the task and during a post‑block discussion period, suggesting less neural alignment between human partners. Notably, despite the AI’s higher objective accuracy, participants showed a greater preference for following human advice over AI advice, and following AI advice actually led to higher decision accuracy, revealing a paradox in advice adoption.