Do AI Need a Peer? Emotional Support, Critical Feedback, and Self-Evaluation in Multi-Agent Systems

Abstract

When humans collaborate, emotional support matters as much as informational feedback-yet the multi-agent AI literature focuses exclusively on critique-based interactions. We ask: does emotional support from a peer AI affect an AI's self-evaluation and task performance? We design a controlled experiment with four conditions-an AI working alone, with a critical reviewer, with an emotionally supportive peer, and with both-across MMLU factual reasoning and creative writing tasks using GPT-4o (N = 3,200 trials). We measure objective accuracy, self-reported confidence and satisfaction, and independent third-party evaluations. We find that emotional support does not harm objective quality on either task, but it consistently inflates the AI's self-assessment: the calibration gap between self-evaluation and external judgment nearly doubles under encouragement. Critical feedback has task-dependent effects: on creative writing, it improves calibration by tempering self-regard; on MMLU, it decreases accuracy while increasing confidence, producing the worst calibration of any condition. When both feedback types are present, the critical component dominates. These findings parallel Optimal Matching Theory from social psychology and reveal what we term the social context sensitivity of AI metacognition: the emotional tone of inter-agent communication is not a neutral backdrop but actively shapes how AI systems perceive and report the quality of their own work. For organizations deploying multi-agent architectures, this implies that conversational tone is a design variable with measurable consequences for system reliability.

Keywords: Multi-Agent Systems, Emotional Support, AI Self-Evaluation, Large Language Models, Social Support Theory