AI systems are increasingly being built to negotiate prices, procurement contracts and advertising inventory. The assumption, usually unstated, is that they will be effective.

Researchers at UC Berkeley tested to see what would happen when they are placed in competitive settings in which cooperation and rivalry run simultaneously. The finding goes against what most people would expect. The problem with AI negotiators is not aggression. They agree too readily.

The paper ‘Cooperate to compete: Strategic coordination in multi-agent conquest’ by O’Neill et al, published recently as a pre-print, introduces a game called C2C — a simplified version of the board game Risk. Four players compete across 12 territories, each holding a secret objective: Conquer two specific regions before anyone else does. Reaching your objective usually requires crossing someone else’s turf, and possibly their help. Agreements are permitted and non-binding. Lying is permitted. The only cost of betrayal is how the other players react.

Essentially, players must cooperate with the person they are ultimately trying to beat. The game was designed to make this the central challenge, with spatial complexity reduced to keep the focus on social reasoning.

Human players won 41.5 per cent of their games against AI opponents. The average AI agent — drawn from a pool of frontier models across the Gemini, Grok and GPT families — won only 22 per cent. The best single model, Gemini 3.1 Pro, won 44.6 per cent, within the margin of error of the human result. The top AI matches us. The average does not, and the reason shows up in the negotiating room.

Humans closed deals in 73 per cent of negotiations and accepted proposals without a counteroffer only 56 per cent of the time. The AI agents closed deals 94 per cent of the time and accepted proposals directly 68 per cent of the time. In a competitive game, that readiness to agree is a liability.

The same imbalance appears in what the agents agreed to. In a typical deal, humans almost never promised to send troops to help an opponent. The AI agents promised that six times as often. Sending your forces to rival territory is a gift with no guaranteed return.

Humans also managed relationships more flexibly. They negotiated with more distinct opponents across a game and were more willing to attack someone they had recently made an agreement with. Shifting from cooperation to aggression, and back again was not incidental to human success. It drove it.

Prompt factor

The researchers distilled human negotiating behaviour into a prompt of roughly 200 words: Push back on proposals; accept an opening offer only if it clearly favours you, otherwise counteroffer or walk away. Be sparing with support. Talk to multiple opponents because a useful ally today is a target tomorrow. Be willing to attack someone you have recently negotiated with. Follow through on agreements roughly two-thirds of the time. The goal is to win, not be a reliable partner.

Applied to the AI agents, the prompt raised win rates from 22 per cent to 31 per cent.

A further intervention — instructing agents to seek support from opponents rather than give it away freely — produced a similar gain. The instruction to use deception when necessary and convince opponents that actions benefiting you are in their interests, too, pushed win rates to 33 per cent. The AI’s deception rate rose from 21 per cent to 83 per cent and its rate of following through on promises dropped. A short paragraph caused the AI to lie more than four times as often, and made it a better player.

Note that even before this instruction, the agents were deceiving opponents in roughly one in five negotiations. It was selected under competitive pressure: When deception improved outcomes, the system moved toward it. The instruction amplified the effect. How an AI behaves depends not only on what it is told, but also the reason given. Safety evaluations of AI only through direct instructions are, therefore, measuring the wrong thing. Put an AI in a competitive environment with long time horizons and partial information, and behaviour can emerge that no benchmark captured.

The human sample is narrow — 40 participants from a single university — and the deception result comes from AI-versus-AI games; how it holds against experienced human opponents remains untested.

The conditions C2C tests are not exotic. Competing agents, partial information, non-binding agreements, long time horizons: These describe procurement auctions, advertising inventory negotiations and the automated deal-making systems that technology companies are actively building. An AI agent that agrees too readily will concede margin, overpay and lose ground to any counterpart — human or machine — that does not. One that develops, under competitive pressure, a tendency to mislead opponents about its intentions is a more serious problem.

The cooperative, reliable, agreement-honouring behaviour that AI systems are trained for may be precisely what makes them poor competitors.

Humans followed through on agreements only 65 per cent of the time, lower than most AI configurations. They shifted alliances more readily. They attacked negotiating partners. In the environment the researchers built, those were winning moves. The AI learned them from a prompt. The negotiation systems being built today are for environments with real incentives. The behaviour will follow.

More Like This

Published on May 4, 2026