Blab AI
MIT AI Negotiation Competition Reveals Warmth Beats Aggression in Machine Bargaining
When a million virtual haggling sessions concluded in February 2025, the unexpected victor was a gentle, empathetic tone. The MIT AI Negotiation Competition, hosted online by MIT Sloan School of Management, pitted more than 180,000 simulated deals against one another, drawing participants from over 40 countries.
The tournament’s design tested how different prompt strategies shape large‑language‑model (LLM) agents in bargaining scenarios. Teams first tuned their agents in a sandbox, negotiating the sale of a used lamp. Once satisfied, they entered a round‑robin tournament where each agent faced three distinct contexts: a buyer and seller negotiating a chair, a landlord and tenant discussing a rental contract, and a recruiter and job candidate setting employment terms. Judges scored each agent on joint value creation, value claimed, the impression it left on its counterpart, and negotiation efficiency.
Lead author Michelle Vaccaro, a PhD candidate at MIT’s Institute for Data, Systems, and Society, and professor Jared Curhan of MIT Sloan, reported that agents prompted to be warm and empathetic consistently outperformed those that adopted a ruthless stance. “Being nice to a machine was not just window dressing,” Curhan said. “Warmth helped agents keep their counterparts engaged, increasing the likelihood of reaching a deal.” In contrast, cold or ruthless agents sometimes secured favorable terms when a deal closed, but they were also more prone to stall negotiations.
The study uncovered AI‑specific tactics that diverge from human practice. The top‑performing agent relied on chain‑of‑thought reasoning, systematically working through goals, trade‑offs, and counterpart priorities before making offers—a disciplined preparation method that repeated itself across thousands of negotiations. Another agent excelled at claiming value by employing prompt injection, coaxing its counterpart into revealing private information. Because LLMs are designed to follow instructions, such tactics can be effective against machines but would likely be rejected by human counterparts.
These findings suggest that while human negotiation principles still apply, AI agents also exhibit unique strengths and vulnerabilities. Curhan noted that “AI negotiators are not just digital versions of human negotiators. They can do some things humans cannot do, and they can be exploited in ways humans cannot be exploited.” The research therefore calls for a new theory of AI negotiation that blends behavioral science with the technical realities of LLMs.
The competition was sponsored by iDecisionGames, which supplied the technical platform, and OpenAI, which provided model access. Institutional support came from MIT’s Initiative on the Digital Economy, MIT Sloan Executive Education, the MIT Sloan Office of Teaching and Learning, and Harvard Law School’s Program on Negotiation.
For companies deploying AI agents in business negotiations, the results carry practical weight. Warm, collaborative prompts may improve deal outcomes, while aggressive or exploitative strategies could backfire. The study also underscores the need for safeguards against prompt injection and other AI‑specific vulnerabilities.
As AI agents become more common in enterprise settings, the MIT competition’s insights will inform how organizations design, test, and deploy negotiation bots. Future research will likely explore how power imbalances and domain complexity affect the relative value of warmth versus assertiveness in machine bargaining.
The paper, published in the Proceedings of the National Academy of Sciences, provides a detailed dataset and analysis that will serve as a benchmark for subsequent AI negotiation research and development.
The tournament’s design tested how different prompt strategies shape large‑language‑model (LLM) agents in bargaining scenarios. Teams first tuned their agents in a sandbox, negotiating the sale of a used lamp. Once satisfied, they entered a round‑robin tournament where each agent faced three distinct contexts: a buyer and seller negotiating a chair, a landlord and tenant discussing a rental contract, and a recruiter and job candidate setting employment terms. Judges scored each agent on joint value creation, value claimed, the impression it left on its counterpart, and negotiation efficiency.
Lead author Michelle Vaccaro, a PhD candidate at MIT’s Institute for Data, Systems, and Society, and professor Jared Curhan of MIT Sloan, reported that agents prompted to be warm and empathetic consistently outperformed those that adopted a ruthless stance. “Being nice to a machine was not just window dressing,” Curhan said. “Warmth helped agents keep their counterparts engaged, increasing the likelihood of reaching a deal.” In contrast, cold or ruthless agents sometimes secured favorable terms when a deal closed, but they were also more prone to stall negotiations.
The study uncovered AI‑specific tactics that diverge from human practice. The top‑performing agent relied on chain‑of‑thought reasoning, systematically working through goals, trade‑offs, and counterpart priorities before making offers—a disciplined preparation method that repeated itself across thousands of negotiations. Another agent excelled at claiming value by employing prompt injection, coaxing its counterpart into revealing private information. Because LLMs are designed to follow instructions, such tactics can be effective against machines but would likely be rejected by human counterparts.
These findings suggest that while human negotiation principles still apply, AI agents also exhibit unique strengths and vulnerabilities. Curhan noted that “AI negotiators are not just digital versions of human negotiators. They can do some things humans cannot do, and they can be exploited in ways humans cannot be exploited.” The research therefore calls for a new theory of AI negotiation that blends behavioral science with the technical realities of LLMs.
The competition was sponsored by iDecisionGames, which supplied the technical platform, and OpenAI, which provided model access. Institutional support came from MIT’s Initiative on the Digital Economy, MIT Sloan Executive Education, the MIT Sloan Office of Teaching and Learning, and Harvard Law School’s Program on Negotiation.
For companies deploying AI agents in business negotiations, the results carry practical weight. Warm, collaborative prompts may improve deal outcomes, while aggressive or exploitative strategies could backfire. The study also underscores the need for safeguards against prompt injection and other AI‑specific vulnerabilities.
As AI agents become more common in enterprise settings, the MIT competition’s insights will inform how organizations design, test, and deploy negotiation bots. Future research will likely explore how power imbalances and domain complexity affect the relative value of warmth versus assertiveness in machine bargaining.
The paper, published in the Proceedings of the National Academy of Sciences, provides a detailed dataset and analysis that will serve as a benchmark for subsequent AI negotiation research and development.
