Apr 16th 2025
IN A MESSY age of grinding wars and multiplying tariffs, negotiators are as busy as the stakes are high. Alliances are shifting and political leaders are adjusting—if not reversing—positions. The resulting tumult is giving even seasoned negotiators trouble keeping up with their superiors back home. Artificial-intelligence (AI) models may be able to lend a hand.
Some such models are already under development. One of the most advanced projects, dubbed Strategic Headwinds, aims to help Western diplomats in talks on Ukraine. Work began during the Biden administration in America, with officials on the White House’s National Security Council (NSC) offering guidance to the Centre for Strategic and International Studies (CSIS), a think-tank in Washington that runs the project. With peace talks under way, CSIS has speeded up its effort. Other outfits are doing similar work.
The CSIS programme is led by a unit called the Futures Lab. This team developed an AI language model using software from Scale AI, a firm based in San Francisco, and unique training data. The lab designed a tabletop strategy game called “Hetman’s Shadow” in which Russia, Ukraine and their allies hammer out deals. Data from 45 experts who played the game were fed into the model. So were media analyses of issues at stake in the Russia-Ukraine war, as well as answers provided by specialists to a questionnaire about the relative values of potential negotiation trade-offs. A database of 374 peace agreements and ceasefires was also poured in.
Thus was born, in late February, the first iteration of the Ukraine-Russia Peace Agreement Simulator. Users enter preferences for outcomes grouped under four rubrics: territory and sovereignty; security arrangements; justice and accountability; and economic conditions. The AI model then cranks out a draft agreement. The software also scores, on a scale of one to ten, the likelihood that each of its components would be satisfactory, negotiable or unacceptable to Russia, Ukraine, America and Europe. The model was provided to government negotiators from those last three territories, but a limited “dashboard” version of the software can be run online by interested members of the public.
The Futures Lab is also designing add-on models for the simulator. Each is a bot trained on studies of, and speeches and writings by, a different political or military leader. To help negotiators work out how China’s president, Xi Jinping, might react to a scenario, for example, an AI alter ego, dubbed “Xibot”, is being developed. The bots also stimulate creativity, says Benjamin Jensen, the lab’s director. His team has already produced three such AI “advisers” that reason in the distinct styles of George Patton, Genghis Khan and Sun Tzu.
Britain’s Foreign Office, meanwhile, is helping fund a more ambitious AI negotiations adviser that is being developed at the University of California, Berkeley. A lab there is training a model exclusively on documents relating to America’s NSC, including minutes of meetings that stretch back to 1951. The idea, says Andrew Reddie, the project’s leader, is to produce a versatile AI adviser for negotiators. The model will produce talking points in a wider range of “voices” than CSIS’s bot advisers.
Demand for such models is high, reckons a negotiator based in London who advises senior government officials in talks on war and peace. When discussions are in full swing behind closed doors, negotiators can lack a way to quickly gauge the opinion of superiors back home, says the adviser (who asked to remain anonymous owing to the sensitive nature of her work). She therefore often needs to pause talks to make contact, which breaks momentum and gives the other side time to regroup. A negotiating team with an AI model that allows it to speed up its tempo in talks, she says, could gain an edge. A good simulator may also be able to flag potential sticking-points early, she adds, as well as helping negotiators see through the eyes of an unsavoury or unrelatable adversary.
Consummate AI diplomats of this sort are still some way off. In tests to identify differences in the negotiation styles of seven AI models, the Futures Lab found that some, including DeepSeek, Gemini and Llama, are particularly “escalatory”. In one scenario Llama opted to use force in a whopping 45% of runs. In other cases, notes Jacquelyn Schneider of Stanford University’s Hoover Institution, a model’s code may be overly conciliatory. Her team compiles data on how various AIs play war games involving negotiation, and briefs congressional staffers on the findings. The “risk-averse” camp includes GPT-4, an OpenAI model Dr Schneider describes as partial to “Obama’s foreign policy”.
Comparative analysis such as Dr Schneider’s could improve future models. So might a new AI project at DARPA, a research agency at the Pentagon. Called CODORD, it aims to convert natural human language about acceptable and unacceptable actions, as well as obligations, into code. That, it is hoped, will help models better hew to a human leader’s intent.
Futures Lab’s next step is to soup up its Ukraine-talks simulator with game theory, which models decision-making by predicting the likely actions of stakeholders based on their goals, motivations and levels of influence. Whereas language models make inferences from existing data, game theory employs deductive reasoning from first principles. Folding it in should give the simulator firmer footing to spot logical errors or unmerited results, says Yasir Atalan, a data scientist at the lab.
One game-theory model to be added is called “Competition in the Shadow of Technology”. Its equations, developed by Futures Lab to increase a country’s negotiating power, calculate the best time for secret military capabilities to be revealed.
The lab also has its eyes on a game-theory model called Predictioneer’s Game which, as a stand-alone system, has an impressive record. Its developer, Bruce Bueno de Mesquita, has used the model to advise clients, including America’s Central Intelligence Agency and the Pentagon, on subjects including nuclear negotiations with Iran and North Korea. A three-day session with the software (and Dr Bueno de Mesquita) can cost $300,000. Though cagey with details, he says he currently discusses the model’s take on “a wide array” of crises with officials at the State Department.
The model has had impressive results. In 2023 Predictioneer’s Game forecast that peace talks on Ukraine would begin in early 2025. In the early days of the conflict in Gaza, the model laid out how Israeli positioning might change if Yahya Sinwar and Ismail Haniyeh, two leaders of Hamas, were ever to exit the picture. That has been borne out by developments following their deaths in 2024, says Dr Bueno de Mesquita. Now 78 years old, he plans to publish the model’s equations for anyone to use.
All this is heady stuff. AI’s potential to reshape security talks, says Rose Gottemoeller, America’s chief negotiator with Russia for New START, a treaty on nuclear arms that took effect in 2011, is “really remarkable”. If the technology catches on, diplomacy may become a field in which AI models reach deals with one another. Whether humans can hold on to a seat at the table is up for discussion. ■
Curious about the world? To enjoy our mind-expanding science coverage, sign up to Simply Science, our weekly subscriber-only newsletter.
Artificial intelligenceScience & technologyComputing
This article appeared in the Science & technology section of the print edition under the headline “AI and the art of negotiations”
Discover stories from this section and more in the list of contents
The Economist today
A daily newsletter with the best of our journalism
Yes, I agree to receive exclusive content, offers and updates to products and services from The Economist Group. I can change these preferences at any time.