When GenAI Negotiates Like Humans: Evidence of the Anchoring Bias in Large Language Models’ Outputs

GenAI-based Large Language Models (LLMs) have revolutionized how people work, communicate, and make decisions. Although most people are aware of GenAI hallucinations, we still rely on GenAI to consult on critical issues. In six studies across three negotiation scenarios, we examine whether two well-known LLMs’ (ChatGPT and Claude) outputs exhibit the anchoring effect, a bias that affects individuals’ counteroffers in negotiations, “anchoring” them to first offers. The LLMs’ outputs exhibited this effect across all six studies. However, neither a perspective-taking prompt about counterpart alternatives (Study 3) nor an LLM-tailored chain-of-thought debiasing prompt (Study 4) eliminated this effect. These findings have important implications for how LLMs are designed and developed, as well as for their growing use as “co-pilots” in human decision-making. While further research across additional models, interventions, and biases is needed, our findings suggest potential risks in treating LLM outputs as bias-free guidance.