AI and the Two Worlds of Negotiation

Abstract Artificial intelligence has already begun to transform negotiation research, teaching, coaching, mediation, competition, and practice. Much of this progress has occurred in what Joel Leibo, drawing on Leonard Savage’s distinction, characterized as “small-world” settings: relatively well-specified environments in which states, payoffs, rules, parties, issues, or evaluation criteria can be sufficiently enumerated to permit precise modeling, simulation, coaching, coding, benchmarking, or optimization. These advances are real and important. AI systems can analyze transcripts, support deliberate practice, generate and evaluate negotiation strategies, facilitate structured dialogue, and permit systematic comparison among competing approaches at scales previously unimaginable. By contrast, “larger world” negotiations are characterized by less well-defined or less-fully-specified versions of these tasks. Yet for negotiation, the small-world / large-world continuum carries a further strategic implication. The most important frontier for AI and negotiation lies not only in building systems that make better moves within relatively fixed “games.” Rather, this wider frontier for AI consists of building systems that help negotiators decide whether the apparent game should be accepted as fixed or be structurally changed in favorable ways by participants or interested third parties who would benefit thereby. Many consequential negotiations unfold toward the larger-world end of this continuum, where the parties, interests, issues, Best Alternatives to Negotiated Agreements (“BATNAs”), action sequences, and process choices are uncertain, dynamic, and subject to purposive change by what we might call “entrepreneurial negotiators.” In these settings, the central challenge is not simply optimization within a well-defined structure, but judgment about the conditions in which the actual and potential structure itself may need to be discovered, challenged, or redesigned in favorable ways. I argue that the next stage of AI-and-negotiation work should be what Raiffa calls “asymmetrically prescriptive.” This means generating optimal advice “against” either 1) known negotiating counterpart(s) or 2) careful descriptions of “nature” as if it were an impersonal other side. Ideally, this prescriptive advice should be theoretically and empirically grounded, practically oriented, and designed to assist human negotiators as they diagnose and shape the game advantageously . Properly guided, AI can help negotiators think strategically and act opportunistically in the face of uncertainty, complexity, and evolving possibilities. Poorly guided, AI may compress negotiation toward largely irrelevant averages, reproduce biases, overfit to narrow scenarios, and/or encourage premature confidence in badly specified games. The task before scholars and practitioners is therefore not merely to ask whether AI can negotiate, but whether AI can help human beings negotiate more wisely when the negotiation itself may not yet have been fully imagined and specified.