Reading Between the Lines: Comparative Accuracy of Large Language Models and Human Negotiators in Predicting Partner Satisfaction

Abstract: Satisfaction is an important outcome that has long-term consequences in negotiation. Although there is ample research on negotiators’ own subjective experience of bargaining, no prior work has directly tested negotiators’ ability to judge partner satisfaction. Across four studies, we examine whether human negotiators can accurately infer their partner’s satisfaction and specifically how their accuracy compares to that of Large Language Models (LLMs). Using chat-based distributive negotiations, we find that negotiators are strikingly poor predictors because they anchor their judgments on their own satisfaction. By contrast, LLMs substantially outperform negotiators and match or exceed third-party human coders in predicting partner satisfaction from the same transcripts. Exploratory linguistic analyses show that negotiators overweight their own utterances, whereas actual partner satisfaction is better predicted from the partner’s language. These findings shed light on a fundamental asymmetry in human versus LLM social inference and suggest new avenues for improving negotiator perception and outcomes.

Keywords: Negotiation, Large Language Models, Satisfaction, Metacognition