Prompted to Feel, Primed to Punish: Large Language Models Use Emotion to Enforce Fairness

Emotions are central to human life, guiding behaviours such as costly punishment of unfairness. Across three studies comprising 429,980 responses from 1,159 human participants and 4,068 large language model (LLM) agents, we tested whether LLMs use emotions to guide altruistic punishment in ways comparable to humans. In Study 1, humans (n = 1,017) and four LLMs (GPT-3.5-turbo-0125, o3-mini, DeepSeek-V3 and DeepSeek-R1) exhibited emotion-behaviour coupling: unfair allocations elicited stronger negative emotions which, in mediation analyses, predicted higher punishment. In Study 2a, prompting reasoning-optimised LLMs (o3-mini and DeepSeek-R1) to self-report their emotional valence and arousal significantly increased punishment, and this causal effect was replicated in humans (Study 2b, n = 142). Alongside these parallels, LLMs were less cost-sensitive and showed stronger emotion-mediated effects than humans, revealing a “hyper-fairness” tendency. These findings provide the first empirical evidence that modern LLMs move beyond emotional knowledge to active emotion utilisation, while highlighting the potential and safeguards for emotionally intelligent AI systems.