Saturday March 8, 2025 – Sunday March 9, 2025
The MIT Samburg Center | 50 Memorial Drive | Cambridge, Massachusetts
Welcome and Introduction
Guhan Subramanian, Harvard Law School and Harvard Business School
Jared Curhan, Massachusetts Institute of Technology
Jonathan Gratch, University of Southern California
On March 8 and 9, 2025, the Program on Negotiation convened leading scholars on AI negotiation to present their cutting-edge research and discuss innovations in the field.
Panel 1: AI Negotiation Competitions
Panel Leader: Jared Curhan (MIT)
Jared Curhan, Massachusetts Institute of Technology
The MIT AI Negotiation Competition
Inspired by Robert Axelrod’s (1984) famous Evolution of Cooperation tournament, the MIT AI Negotiation Competition enlisted negotiation experts, students, and faculty from over 50 countries to prompt large language models (LLMs) to engage in a wide variety of integrative and distributive negotiation tasks. Their challenge was to maximize value claiming (proportion of resources captured), value creation (joint benefit), and subjective value (counterpart satisfaction). A round-robin design involving thousands of negotiations yielded a rich and multifaceted dataset. Preliminary findings suggest that, as in Axelrod’s original tournaments, “nice” strategies play a surprisingly pivotal role in objective and subjective negotiation outcomes. This presentation will share key results, highlighting innovative prompts developed by the competition’s winners and examining broader theoretical implications for negotiation and AI.
Joel Leibo, Google DeepMind and Kings College
The Melting Pot Contest
Multi-agent artificial intelligence research promises a path to develop intelligent technologies that are more human-like and more human-compatible than those produced by “solipsistic” approaches, which do not consider interactions between agents. Melting Pot is a research tool developed to facilitate work on multi-agent artificial intelligence and provides an evaluation protocol that measures generalization to novel social partners in a set of canonical test scenarios. Each scenario pairs a physical environment (a “substrate”) with a reference set of co-players (a “background population”), to create a social situation with substantial interdependence between the individuals involved. For instance, some scenarios were inspired by institutional-economics-based accounts of natural resource management and public-good-provision dilemmas. Others were inspired by considerations from evolutionary biology, game theory, and artificial life. Melting Pot aims to cover a maximally diverse set of interdependencies and incentives. It includes the commonly-studied extreme cases of perfectly-competitive (zero-sum) motivations and perfectly-cooperative (shared-reward) motivations, but does not stop with them. As in real-life, a clear majority of scenarios in Melting Pot have mixed incentives. They are neither purely competitive nor purely cooperative and thus demand successful agents be able to navigate the resulting ambiguity.
Johnathan Mell, University of Central Florida
The Automated Negotiating Agents Competition
Dr. Johnathan Mell will join the panel to discuss his five-year leadership of the Human Agent League at the Automated Negotiating Agents Competition (ANAC) from 2017 to 2021. During this period, Dr. Mell developed and maintained the IAGO platform, an online negotiation agent system that served as the competition platform and contributed to advancing the field of human agent negotiation. ANAC was a sponsored competition held at major conferences like IJCAI and AAMAS, attracting contributions from numerous academic and private industry groups. This international collaboration not only enhanced the state of the art in negotiation systems but also fostered new cognitive theories and practical insights into human-focused negotiating agents. Dr. Mell will discuss the significant advancements achieved during the competition, the complexities of sustaining a platform that supported cutting-edge research in artificial intelligence and cognitive science, and the potential future of IAGO 2.0 and subsequent generations of negotiating agent competitions.
Discussant
Robert Axelrod, University of Michigan
Panel 2: AI as a Researcher
Panel Leader: Ray Friedman (Vanderbilt)
Ray Friedman, Vanderbilt University
Developing a Large Language Model for Coding Negotiation Transcripts
Coding negotiation transcripts is a costly and laborious task that is needed for many research projects. We developed AI coding models that were trained (using in-context learning) on a corpus of previously-human-coded transcripts (Alsani et al 2014; Brett and Nandkeolyar [unpublished]). Five transcripts were used to train the Anthropic LLMs and the model was tested with a different set of 30-60 transcripts. We run each model five times, reporting model’s level of consistency, as well as the code assigned. Initial levels of human-model match were 73% for Model 1 and 75% for Model 2. We then conducted “mismatch analysis” with new human coders assessing whether they judged the initial human coders or the model to be more accurate, raising significantly the expected accuracy level of the models. The models are available at https://www.ainegotiationlab-vanderbilt.com/ and grants to cover the Anthropic charges can be requested at https://www.negotiationandteamresources.com/automated-coding-of-negotiation-transcripts/
Emily Hu, University of Pennsylvania
AI as Explorer: Quantifying Conversations with Natural Language Processing
Conflict and negotiation are rich with language data. Individuals use words to introspect and reason about their values; groups debate, make offers, and reach verbal agreements. These dynamics, though incredibly rich, are difficult to quantify; which attributes in a conversation are worth measuring? And how should any given attribute be operationalized? In this talk, I show how a suite of natural language processing techniques can be used to explore unstructured language in a structured manner. By unifying two disciplines — theories from communication and methods from computer science — I will seek to demystify the process of analyzing open-ended communication data. I introduce the Team Communication Toolkit, a “one-stop shop” Python package that extracts more than 160 communication features from a transcribed conversation. I will then highlight design details, key measures, and considerations for different settings. Finally, I will use the toolkit to explore a dataset of 376 groups playing real-time public goods games, using conversation features to predict cooperative behavior. This application demonstrates the promise of natural language processing to generate hypotheses and obtain theoretical insights from complex social phenomena.
Gale Lucas, University of Southern California
Let’s Negotiate! Using AI as a Partner in Negotiation Research
Agents have been established as useful confederates for research in social sciences, including on the topic of negotiation. Standing in place of human confederates, there is a greater degree of experimental control and thus internal validity. Compared to scripted opponents in research, though, they allow for an interactive, contingent negotiation with players. In this talk, I argue that this tension between control and interactivity also exists among types of agents: agents where users negotiate through drop-down menus afford a great degree of control, but are not as interactive as natural language (NL) agents that respond to free-text responses from players. Agents’ response quality, however, needs to be improved in order to retain experimental control. If possible, this could make NL agents the best of both worlds. The talk ends with a discussion around efforts to address control and response quality, accordingly.
Zhivar Sourati, University of Southern California
LLMs and the Erosion of Human Variance
LLMs offer exciting possibilities for psychological research, from modeling aspects of human psychology to automating tasks like text annotation. However, their uncritical adoption poses significant risks. In this talk, I explore the complex interplay between LLMs and psychological research, addressing both their potential and their perils. I examine how LLMs can inadvertently homogenize language, eroding the rich linguistic diversity crucial for understanding individual identities, psychological states, and social contexts. Furthermore, I argue that using LLMs as human substitutes or direct models of human thought presents distinct challenges. I specifically focus on LLMs’ difficulty in generating the natural variance inherent in human data, which further obscures valuable insights into individual differences. Specifically, I demonstrate that LLMs fail to produce much variance in their answers to psychology survey questions—even those pertaining to topics like moral judgment, where there is no objectively “correct” answer and where human responses would naturally reflect a diversity of thought. These findings underscore the critical need for caution and methodological rigor when using LLMs in psychological research, emphasizing that they should be viewed as tools to augment, not replace, human-centered research approaches.
Panel 3: AI as a Negotiator and Mediator
Panel Leader: Jonathan Gratch (USC)
Jonathan Gratch, University of Southern California
A Brief History of AI Negotiators
With the recent rise of language models like GPT, there has been an explosion of interest in AI as a negotiator. Yet negotiation has been a major focus of AI research since the early 1990s. The field has developed sophisticated techniques to address key challenges in deploying AI negotiators in real-world settings, including aligning AI with user goals, developing provably efficient solutions, incorporating human factors such as emotion and trust, and highlighting ethical concerns when organizations delegate negotiations to AI. These challenges and solutions are particularly relevant for this new generation of language-based approaches. I will briefly review these key challenges and solutions and thereby set the stage for the other panel speakers.
Michelle Vaccaro, Massachusetts Institute of Technology
Smooth-Talking Bots: AI Negotiators Make Better Impressions
This study examines how large language model (LLM) chatbots perform in negotiations compared to humans and explores whether manipulating chatbot “personality” (warmth and dominance) affects negotiation outcomes. Building on theories of algorithmic aversion and algorithmic appreciation, we conducted an integrative experiment in which human participants negotiated with human, chatbot, or Wizard-of-Oz (human-controlled) counterparts. Our design spanned four “spaces”: (1) actor (human or chatbot), (2) personality (systematically varied warmth-dominance combinations), (3) task (distributive or integrative bargaining), and (4) role (e.g., buyer vs. seller). Preliminary results suggest that unprompted chatbots outperform human negotiators in subjective value (i.e., counterpart satisfaction) but underperform on objective outcomes in both distributive and integrative contexts. Additionally, prompting chatbots to exhibit warmth increases subjective value at the expense of objective outcomes, whereas prompting them to display dominance strengthens objective outcomes but reduces subjective value. However, when warmth and dominance are combined, chatbots achieve both higher subjective and objective outcomes than unprompted chatbots. By demonstrating that AI negotiators can improve user experience without compromising deal terms, this research offers practical implications for organizations considering the use of LLM-enabled negotiation systems. Future work will involve larger samples to deepen our understanding of how chatbot personality interacts with individual human differences—such as trust in AI and cultural norms—to shape negotiation dynamics and outcomes.
Mike Lewis, Meta AI Research
What We Learned from Teaching AI to Negotiate
Negotiation has long been an interesting problem to AI researchers, as it relies on building agents that can strategically use language to achieve their own goals, and research on the problem has motivated the development of new techniques. I will describe the first AI system trained end-to-end to negotiate, featuring early uses of planning and end-to-end reinforcement learning to optimize dialogue systems. I will also discuss how we designed an agent game of Diplomacy, in which players must establish trust and coordinate in a competitive environment. Our Cicero bot uses an interface between a large language model and a symbolic planner to perform at the level of strong human players.
Min Kyung Lee, University of Texas at Austin
Opportunities and Challenges in AI Mediators
As AI systems become increasingly embedded in decision-making and social processes, they are taking on the role of mediators—moderating group interactions, facilitating knowledge exchange, and representing individuals in collective decision-making. In this talk, I will explore the opportunities and challenges of AI mediators, drawing from past and ongoing case studies in resource allocation, knowledge sharing, and deliberation.
Panel 4: AI as a Coach
Panel Leader: Michael Morris (Columbia)
Michael Morris, Columbia University
ACE: A LLM-based Negotiation Coaching System
The growing prominence of LLMs has led to an increase in the development of AI tutoring systems. These systems are crucial in providing underrepresented populations with improved access to valuable education. One important area of education that is unavailable to many learners is strategic bargaining related to negotiation. To address this, we develop a LLM-based Assistant for Coaching nEgotiation (ACE). ACE not only serves as a negotiation partner for users but also provides them with targeted feedback for improvement. To build our system, we collect a dataset of negotiation transcripts between MBA students. These transcripts come from trained negotiators and emulate realistic bargaining scenarios. We use the dataset, along with expert consultations, to design an annotation scheme for detecting negotiation mistakes. ACE employs this scheme to identify mistakes and provide targeted feedback to users. To test the effectiveness of ACE-generated feedback, we conducted a user experiment with two consecutive trials of negotiation and found that it improves negotiation performances significantly compared to a system that doesn’t provide feedback and one which uses an alternative method of providing feedback.
Harang Ju, Massachusetts Institute of Technology
Does AI Coaching Improve Negotiation Outcomes? Evaluating Warmth, Dominance, and Individual Traits in Negotiation
Artificial intelligence (AI) tools leveraging large language models (LLMs) have enhanced task performance across domains of knowledge work. Yet, their impact on negotiation—a critical component of collaborative work—remains unclear. This pre-registered online experiment investigates how real-time AI coaching, delivered through distinct communication styles (warmth- or dominance-oriented prompts), shapes negotiation outcomes. We hypothesize that AI assistance improves both objective (e.g., deal value) and subjective (e.g., perceived fairness) outcomes, with effectiveness contingent on alignment between coaching style and individual receptivity. Participants were paired and randomized into conditions including no AI, AI with fluency-focused prompts, and AI with combined warmth or dominance strategies. During negotiations, participants received AI-generated suggestions for chat messages. The study evaluates (1) whether AI coaching enhances outcomes, (2) which coaching styles (warmth, dominance, fluency) are most effective, and (3) which individuals benefit most from specific styles. This work advances understanding of how AI coaching can enhance negotiation dynamics in socially complex contexts.
Peter Carnevale, University of Southern California
AI Knows a Deal When It Sees One
Classification of negotiated agreements is important for understanding the basic structure of agreement and for coaching people on what is possible in negotiation. An important step is answering this question: Can people, and AI, reliably recognize and classify agreements? If yes, it suggests new avenues for systematic investigation on agreement and coaching negotiators to consider possibilities (cf. Lax & Sebenius’s, 2006, notion of “deal design”). Pruitt developed a 5-category classification scheme (1981) and Carnevale (2006) expanded it to 8, with a new category called “Superordinate” that derived from Sherif’s (1958) notion of the “superordinate goal” wherein disputing parties forgo an immediate interest in lieu of a common goal, that is, a replaced interest. A Superordinate agreement is one where the parties drop their initial interests in favor of new interests that emerge. A coding manual for the 8-agreement taxonomy was developed that includes definitions and examples of agreements of each type. The OpenAI ChatGPT4o was trained on the agreement taxonomy. Can people, and AI, reliably recognize and classify agreements? Yes. As an example, see Figure 1 that shows an instance of an agreement that had not previously been shown to GPT, and GPT’s classification. It is worth noting that a large sample of human judges concurred that the agreement is Expand. The human-AI concurrence across instances of the 8 different agreement categories was remarkable. As a further test of GPT’s ability to classify agreement, it was asked to read the Mayflower Compact (Figure 2). Then it was asked to indicate which if any of the 8 agreement types describes the Mayflower Compact. The Mayflower Compact, signed in 1620, is a historic agreement that established self-governance in the Plymouth Colony. It is one of the first formal agreements made by European settlers in the Western Hemisphere. The 41 signers, having landed unexpectedly outside their intended destination, faced internal group conflict. We expected it to be difficult to classify. But GPT4o sees the Mayflower Compact as Superordination, one of the eight basic forms of agreement. See Figure 3. Again, human judges concurred.
Amy Schmitz, The Ohio State University
AI as the Fourth Party in Empowering Self-Represented Litigants
JusticeTech is an innovative initiative that explores how AI can act as the “fourth party” to help self-represented litigants (SRLs) navigate negotiations and mediation in vital areas impacting Access to Justice (A2J).
The collaborative efforts between the Moritz College of Law and the Computer Science and Engineering (CSE) department at Ohio State University are unique, involving researchers from both departments and Public Affairs. JusticeTech focuses on developing AI-enhanced Online Dispute Resolution (ODR) systems to expand A2J, particularly for SRLs in eviction and truancy cases. Ethical AI design is a cornerstone, ensuring technologies mitigate bias and safeguard individual rights.
Key projects include the Eviction Project, which involves empirical research and focus groups to understand the legal landscape of evictions, supported by nearly $1.6 million in grants. The Truancy Project addresses truancy issues in Ohio through online mediation and AI-enhanced assistance, with ongoing feedback from community partners and expert mentors. JusticeTech’s interdisciplinary approach involves project-based learning, pro-bono assistance, and robust research to provide insights into how technology can assist SRLs.
Panel 5: AI as a Teacher
Panel Leader: Lawrence Susskind (MIT)
Lawrence Susskind (MIT)
Panel Leader and Discussant
Laura Wang, Massachusetts Institute of Technology
The Negotiation Skills Assessment: Developing and Validating an AI-Powered Measure of Negotiation Proficiency
This study introduces the Negotiation Skills Assessment (NSA), an innovative framework and tool that integrates traditional assessment techniques with AI-driven negotiation simulations to evaluate core negotiation skills. By leveraging advancements in AI, the NSA addresses persistent challenges in negotiation assessment, such as biases in self-reported evaluations and the overemphasis on objective outcomes. A pilot implementation revealed significant improvements in students’ negotiation skills following a three-day negotiation course, highlighting the effectiveness of the NSA. These findings demonstrate the potential of combining AI technologies with established evaluation methods to develop scalable, accurate, and engaging tools for assessing and enhancing negotiation abilities.
Emmanuel Dorley, University of Florida
Virtual Agents for Personalized Negotiation Training
Advancements in AI have made it possible to develop agents capable of engaging in negotiation, offering new opportunities for personalized training. Despite its importance, negotiation remains underemphasized in most disciplines and, specifically, in engineering education. In this talk, I will present our research on using AI agents as role-playing partners to help students learn how to negotiate. The approach allows students to practice learned negotiation strategies with an agent and receive personalized feedback on their performance. I will also discuss our ongoing work to enhance these simulations, making negotiation training more accessible, scalable, and personalized.
Samuel “Mooly” Dinnar, Massachusetts Institute of Technology
Negotiation Coaching Bots and Backtable Bots: Using GenAI to Improve Human-to-Human Interactions in Multiparty Negotiation Instruction
This session describes the Negotiation Coaching Bots using GenAI that were developed for teaching in two muti-party negotiation courses at MIT (Spring 2024) and the resulting papers about the experience. Three kinds of bots were created for various stages: a PREPARATION coaching bot for each role-player before entering the actual human-human multi-party interaction in an upcoming specific simulation; a BACKTABLE bot that allows learners to interact with (imaginary) backtable stakeholders from the assigned negotiation partners; and a DEBRIEF bot that coaches individuals to reflect deeply after their negotiation. We share our process of Seven Steps to Building GenAI Negotiation Bots, paying specific attention to the important Four Aspects of Prompt Design, and how we provided the students a choice about their coach’s personality style. We’ll present data from the students’ learning steps and reactions to being coached with bots in 2 different 6-person role-plays, emphasizing the importance of context and asking-questions to productive coaching, and to developing the students’ Personal Theory of Practice. We’ll conclude with ideas and recommendations regarding the further possible development of Negotiation Coaching Bots, and how they can be used in a wide range of teaching and training settings.
Panel 6: AI in the Field
Panel Leader: Jeanne Brett (Northwestern)
Jeanne Brett, Northwestern University
NegotiAge: Development and Pilot Testing of an Artificial Intelligence-Based Family Caregiver Negotiation Program
Over 44 million family caregivers provide unpaid assistance and support to older adults. Most are unprepared for their role and often provide care with little or no support. Caregiving is associated with decline in physical health, mental well-being, and overall quality of life. The purpose of the research was to design and evaluate conflict management training for family caregivers with the goal of empowering them to negotiate conflict situations more effectively. Researchers transformed IAGO, an LLM model designed for human-agent negotiations into NegotiAge. The AI agent in NegotiAge uses Ury, Brett, and Goldberg’s three approaches to resolving disputes: Interests, Rights, and Power (IRP) to negotiate as an older adult, a sibling, and a physician. A total of 124 caregivers were randomly assigned to complete one to four negotiation exercises. All had access to instructional videos explaining IRP, a planning document, and negotiation tips. Results showed that caregivers’ knowledge of IRP improved with training and that knowledge was sustained over three months. At the three-month follow-up, 74% of participants reported using the training in their caregiving role; 42% reported applying it in other areas of their lives. Caregivers reported shifting their conflict resolution strategies away from power dynamics toward shared interests. This research demonstrates the feasibility and effectiveness of training caregivers to negotiate the resolution of conflict.
Zilin Ma, Harvard University
“ChatGPT, Don’t Tell Me What to Do”: Designing AI for Context Analysis in Humanitarian Frontline Negotiations
Frontline humanitarian negotiators are increasingly exploring ways to use AI tools in their workflows. However, current AI-tools in negotiation primarily focus on outcomes, neglecting crucial aspects of the negotiation process. Through iterative co-design with experienced frontline negotiators (n=32), we found that flexible tools that enable contextualizing cases and exploring options (with associated risks) are more effective than those providing direct recommendations of negotiation strategies. Surprisingly, negotiators demonstrated tolerance for occasional hallucinations and biases of AI. Our findings suggest that the design of AI-assisted negotiation tools should build on practitioners’ existing practices, such as weighing different compromises and validating information with peers. This approach leverages negotiators’ expertise while enhancing their decision-making capabilities. We call for technologists to learn from and collaborate closely with frontline negotiators, applying these insights to future AI designs and jointly developing professional guidelines for AI use in humanitarian negotiations.
Michiel Bakker, Massachusetts Institute of Technology
The Habermas Machine: AI Can Help Humans Find Common Ground in Democratic Deliberation
Finding agreement through a free exchange of views is often difficult. Collective deliberation can be slow, difficult to scale, and unequally attentive to different voices. In this study, we built “The Habermas Machine”, an LLM-based system to mediate human deliberation. Using participants’ personal opinions and critiques, the AI mediator iteratively generates and refines statements that express common ground among the group on social or political issues. Participants (N = 5734) preferred AI-generated statements to those written by human mediators, rating them as more informative, clear, and unbiased. Discussants often updated their views after the deliberation, converging on a shared perspective. Text embeddings revealed that successful group statements incorporated dissenting voices while respecting the majority position. These findings were replicated in a virtual citizens’ assembly involving a demographically representative sample of the UK population.
Comments are closed.