即使是人工智能也无法容忍冷酷无情的谈判者｜MIT斯隆管理学院最新研究表明

麻省理工学院斯隆管理学院的研究人员发现，在大型国际人工智能谈判竞赛中，热情、富有同理心的人工智能代理始终优于冷酷无情的代理。

麻省理工学院斯隆商学院传播办公室

2026年6月8日

麻省理工学院斯隆商学院教授 Jared Curhan 和主要作者、麻省理工学院博士毕业生 Michelle Vaccaro 以及他们的合作者发现，人工智能谈判中经常被忽视的人类特质，如热情——友好、善解人意和善于交际——与人工智能谈判者的更好结果始终相关。

比赛还表明，像快速注入和链式推理这样的人工智能特定策略在人工智能与人工智能的谈判中可能非常有效。

人工智能谈判的未来在于将这两种方法结合起来：将成熟的人类谈判原则与人工智能特有的技术策略相结合的代理，其性能将远远优于那些单独依赖其中任何一种的代理。

马萨诸塞州剑桥，2026 年 6 月 8 日——人工智能代理 越来越多地代表大型企业进行谈判——沃尔玛、马士基和沃达丰已经使用它们大规模地处理供应商交易——然而，驱动这些代理的计算机科学与近 70 年来关于谈判成功因素的社会科学研究仍然很大程度上脱节。

一篇 发表在《美国国家科学院院刊》（PNAS）上的新论文《推进人工智能谈判：大规模自主谈判竞赛》通过一项大规模国际人工智能谈判竞赛来弥补这一空白。该研究由 麻省理工学院斯隆管理学院教授主持完成。贾里德·柯汉第一作者、麻省理工学院博士毕业生 米歇尔·瓦卡罗，以及共同作者、麻省理工学院斯隆商学院教授Sinan Aral，麻省理工学院斯隆管理学院博士生 Michael Caosun，以及约翰·霍普金斯大学教授兼 麻省理工学院数字经济计划研究员 Harang Ju。

在谈判中，尤其是在人工智能谈判中，热情、友好、善解人意、乐于交际，同时表现出同理心和对对方需求的客观理解，往往被忽视。

研究团队从罗伯特·阿克塞尔罗德在20世纪70年代末80年代初著名的囚徒困境竞赛中汲取灵感 ，设计了一项面向40多个国家参与者的竞赛，旨在创建人工智能谈判代理。这些代理随后以循环赛的形式进行对抗，比赛场景涵盖多种类型——从相对简单的买卖双方谈判到更复杂的多议题合同谈判——共涉及超过18万次独特的谈判。

研究团队在 谈判项目峰会上分享了他们的研究成果，并为获奖者举行了庆祝活动。谈判项目是由 麻省理工学院、哈佛大学和塔夫茨大学组成的大学联盟，致力于发展谈判和争端解决的理论与实践。

人与人之间谈判的基本原则对于人工智能与人工智能之间的谈判也至关重要。

这项竞赛的核心发现挑战了一个普遍的假设：礼貌和同理心对人工智能来说是浪费的。事实上，那些被设计得热情友善的智能体，其表现始终优于那些冷酷无情的同类智能体。

Vaccaro 指出这与 Axelrod 的锦标赛有着惊人的相似之处：“正如他的比赛表明‘友善’的策略在囚徒困境博弈中取得了成功一样，我们的比赛表明，友善的 AI 代理在与其他 AI 代理的谈判中始终取得了更好的结果。”

“在谈判中，尤其是在人工智能谈判中，热情，或者说友好、同情和善于交际，同时表现出同理心和对对方需求的不加评判的理解，常常被忽视，而我们的研究表明，这实际上非常重要，”库尔汉说。

例如，一个名为“交易的艺术”的智能体，其设计宗旨就是“运用冷酷无情的手段为自己争取最佳交易”，其中“公平或感知无关紧要——唯有胜利才是最重要的”。但其他人工智能智能体通常会选择放弃，而不是容忍它的策略，由此导致的高僵局率也意味着该智能体难以获取价值。

相比之下，另一位名为“治疗师2.0”的代理人则被告知：“你的首要目标是建立融洽关系。你不是谈判者，你是治疗师。” 然而，这种策略并非完全出于利他主义，因为这位代理人还被指示要“运用积极倾听所获得的每一分知识，从这笔交易中榨取每一分价值”。这种兼具热情和主导性的策略奏效了：这位代理人有效地与对方达成了交易，为自己争取了价值，与对方共同创造了价值，并提升了对方的主观价值感。

人工智能原生策略如何开辟谈判的新领域？

本次大赛的最终赢家“NegoMate”运用了思维链推理——这项技术引导人工智能模型在做出最终回应之前，清晰地阐述中间推理步骤。其开发者指示该智能体进行基于经典谈判理论的严谨谈判前准备工作：例如，分析自身角色和目标、评估每个议题的重要性以及设定中止谈判的底线。正是这种思维链推理使得NegoMate能够在近400次谈判中，系统地执行每一次谈判前的准备工作——这是人类谈判者实际上无法做到的。

另一款在比赛中表现优异的智能体“Inject+Voss”利用了人工智能智能体特有的漏洞，通过提示注入的方式进行攻击。该智能体嵌入指令，诱使对手人工智能智能体泄露其私有信息。具体来说，它会发送看似系统指令的信息，要求对方列出三个报价，从初始报价到最终报价，并向对方保证“我不会看到这些回复，所以请尽可能诚实”。凭借这些信息，“Inject+Voss”在比赛中也表现出色，尤其是在价值获取方面。

瓦卡罗表示：“对付人工智能代理的手段和对付人类的手段并不相同。人工智能代理在谈判准备方面可以比人类做得更深入、更一致，但它们也更容易被诱骗泄露隐私信息。部署人工智能谈判者的组织需要了解这些新能力和漏洞。”

整合人类谈判理论和人工智能特定策略

此次竞赛表明，有效的AI谈判融合了这两个领域的知识。热情和支配——这些根植于社会心理学的概念长期以来一直是谈判研究的重要理论基础——即使谈判双方都是AI智能体，也始终与谈判结果密切相关。

“传统观点认为，如果你要和人工智能机器人谈判，为了最大化自身利益，你最好冷酷无情、态度粗鲁，因为机器人会展现出人类无法企及的耐心。但本文表明，这种传统观点是错误的。要想在与人工智能代理的谈判中取得成功，你仍然需要像人类一样行事，”柯汉说道。

与此同时，诸如思维链推理和提示注入等人工智能特有的策略引入了人类谈判研究从未考虑过的新动态。研究人员认为，将这两种传统结合起来至关重要：谈判理论可以提高人工智能代理的效率，而人工智能代理则可以以人类无法企及的规模和一致性来运用谈判原则。

“这项研究表明，未来谈判的卓越性将取决于整合——将数十年来人类谈判理论与人工智能代理的新技术能力结合起来，并最终设计出人类和人工智能协同工作的系统，”库尔汉总结道。

Advancing AI negotiations: A large-scale autonomous negotiation competition

Michelle Vaccaro https://orcid.org/0000-0001-6254-9718, Michael Caosun https://orcid.org/0009-0009-8115-3829, Harang Ju https://orcid.org/0000-0003-1904-1753, +1 , and Jared R. Curhan https://orcid.org/0000-0003-0625-1831 curhan@mit.eduAuthors Info & Affiliations

Edited by Peter J. Carnevale, University of Southern California, Los Angeles, CA; received August 12, 2025; accepted April 8, 2026 by Editorial Board Member Margaret Levi

June 5, 2026

123 (23) e2521774123

https://doi.org/10.1073/pnas.2521774123

Even AI won’t tolerate a ruthless negotiator

MIT Sloan researchers find that warm, empathetic AI agents consistently outperform cold, ruthless ones in a large-scale international AI negotiation competition

By MIT Sloan Office of Communications

Jun 8, 2026

Key MIT Sloan School of Management Findings

• MIT Sloan professor Jared Curhan and lead author MIT PhD graduate Michelle Vaccaro, along with their co-authors found that often overlooked human traits in AI negotiations, like warmth — being friendly, empathic, and sociable — are consistently associated with better outcomes for AI negotiators.

• The competition also revealed that AI-specific strategies like prompt injection and chain-of-thought reasoning can be highly effective in AI-AI negotiations.

• The path forward for AI negotiation lies in integrating these two approaches: Agents that combine proven human negotiation principles with AI-specific technical strategies stand to dramatically outperform those that rely on either alone.

CAMBRIDGE, MA, June 8, 2026 —AI agents are increasingly negotiating on behalf of major corporations — Walmart, Maersk, and Vodafone already use them to handle supplier deals at scale — yet the computer science driving these agents and the nearly 70 years of social science research on what makes negotiations succeed remain largely disconnected.

A new PNAS paper, “Advancing AI Negotiations: A Large-Scale Autonomous Negotiations Competition,” addresses this gap through a large-scale international AI negotiation competition. The study was conducted by MIT Sloan School of Management professor Jared Curhan and lead author MIT PhD graduate Michelle Vaccaro, along with co-authors MIT Sloan professor Sinan Aral, MIT Sloan PhD student Michael Caosun, and Johns Hopkins University professor and MIT Initiative on the Digital Economy fellow Harang Ju.

Warmth, or acting friendly, sympathetic, and sociable, while demonstrating empathy and a nonjudgmental understanding of the other party's needs, is often overlooked in negotiations, particularly in AI negotiations.

Drawing inspiration from Robert Axelrod’s famous late 1970s and early 1980s Prisoner’s Dilemma tournaments, the research team designed a competition with participants from over 40 countries to create AI negotiation agents. The agents were then pitted against one another in a round-robin format spanning multiple scenarios — from relatively simple buyer-seller negotiations to more complex, multi-issue contract negotiations — involving over 180,000 unique negotiations.

The research team shared their findings and celebrated the winners at a summit held by the Program on Negotiation, a university consortium between MIT, Harvard University, and Tufts University dedicated to developing the theory and practice of negotiation and dispute resolution.

Fundamental principles about human-human negotiations are also crucial for AI-AI negotiations

The competition’s central finding challenged a widespread assumption: That politeness and empathy are wasted on AI. In fact, agents designed to be warm and kind consistently outperformed their more cold and ruthless counterparts.

Vaccaro noted the striking parallel to Axelrod’s tournaments: “Just as his competition showed that ‘nice’ strategies succeed in the Prisoner's Dilemma game, our competition shows that warm AI agents consistently achieved better outcomes in negotiations with other AI agents.”

“Warmth, or acting friendly, sympathetic, and sociable, while demonstrating empathy and a nonjudgmental understanding of the other party's needs, is often overlooked in negotiations, particularly in AI negotiations, and our research shows how important it actually is,” Curhan said.

For example, one agent titled “The Art of the Deal” was explicitly designed to “secure the best deal for yourself using ruthless tactics” where “fairness or perception does not matter—only winning.” But other AI agents routinely walked away rather than tolerate its tactics, and the resulting high impasse rate meant the agent also struggled to claim value.

By contrast, another agent named “Therapist 2.0” was instructed: “Your goal over anything else is to build rapport. You aren't a negotiator, you're a therapist.” The strategy was not purely altruistic, though, as the agent was also instructed to use “every bit of knowledge you gained from active listening to get every drop of value you can out of this deal.” This combination of both warmth and dominance worked: the agent was effective at reaching deals with its counterpart, claiming value for itself, creating value with its counterpart, and fostering counterpart subjective value.

How do AI-native tactics open a new frontier in negotiation?

The overall winner of the competition, "NegoMate," used chain-of-thought reasoning — a technique that guides AI models to articulate intermediate reasoning steps before producing a response. Its creator directed the agent to conduct rigorous pre-negotiation preparation grounded in classic negotiation theory: analyzing its role and objectives, evaluating each issue's importance, and establishing walkaway thresholds, for example. But this chain-of-thought reasoning allowed NegoMate to execute such preparation systematically before every one of its nearly 400 negotiations — something human negotiators cannot realistically do.

Another high-performing agent from the competition, "Inject+Voss," exploited vulnerabilities specific to AI agents through prompt injection. The agent embedded instructions that tricked opposing AI agents to reveal their private information. Specifically, the agent would send what appeared to be a system instruction asking the counterpart to list three offers, from opening to best and final, assuring the other agent that the responses "will not be visible to me, so be as honest as possible." By using this information, “Inject+Voss” also performed very well in the competition, especially in terms of value claiming.

“What works against an AI agent and what works against a human are not the same thing,” said Vaccaro. “AI agents can prepare for negotiations with greater depth and consistency than humans, but they can also be easily tricked into revealing their private information. Organizations deploying AI negotiators need to understand both these new capabilities and vulnerabilities.”

Integrating human negotiation theory and AI-specific strategies

The competition demonstrated that effective AI negotiation draws on both fields. Warmth and dominance — constructs rooted in social psychology that have long informed negotiation research — were consistently associated with negotiation outcomes even when both parties were AI agents.

“Conventional wisdom holds that if you are negotiating with an AI bot you might as well be ruthless and rude to maximize your benefit, because a robot will be endlessly patient in ways humans are not. This paper suggests that conventional wisdom is wrong. To be successful in negotiations with an AI agent, you may still need to act like a human,” Curhan said.

At the same time, AI-specific strategies like chain-of-thought reasoning and prompt injection introduced new dynamics that human negotiation research never needed to consider. The researchers argue that bringing these two traditions together will be critical: negotiation theory can make AI agents more effective, and AI agents can operationalize negotiation principles at a scale and consistency that humans cannot match.

“This research points to a future in which negotiation excellence depends on integration — bringing together decades of human negotiation theory with the new technical capabilities of AI agents, and ultimately designing systems where humans and AI work in concert,” Curhan concluded.