At our 25-year anniversary at TWS Partners, AI was, unsurprisingly, one of the dominant topics. One point from those discussions has stayed with me: AI and game theory are complementary, and together they will make things possible that neither achieves alone. This article is my attempt to work out what that means, in both directions and for the people who do this for a living.

We usually discuss AI in terms of the tasks it will automate. Research, analysis, coding, presentations, decision support. I think the more interesting question is what these two do for each other.

---

Part one: what AI does for game theory

More decisions become games

Two types of decisions look alike and are not.

Optimizing a production network, forecasting demand or finding the fastest delivery route can be very complicated. But the environment does not react strategically to my decision. That is an optimization problem.

Negotiating with a supplier is different. So is setting a price when competitors can respond, committing capacity when a rival may expand, or building a transfer strategy in football while clubs, players and agents all anticipate each other. Here my best decision depends on what someone else will do because of my decision. That is a game.

A surprising amount of that decision-making is still intuitive. A buyer with twenty years of experience says: “I don’t think the supplier will walk away.” A salesperson assumes the competitor won’t follow the price move. A negotiator pushes once more because “they have room”. Sometimes those intuitions are excellent. Sometimes they aren’t.

Either way, those people will increasingly have an AI sitting next to them. And so will the other side. The buyer uses AI to anticipate the supplier. The supplier uses AI to anticipate the buyer. Competitors do the same to each other. Eventually, autonomous agents may make parts of these decisions themselves. At that point AI is no longer helping us make better individual decisions. AI systems are playing games against other AI systems.

We can finally model the games we used to simplify away

This is the part I find most consequential, and it gets the least attention.

For as long as I have done this work, the most common objection to applied game theory has not been that it is wrong. It has been that reality has too many players, too many moves and too much that changes along the way. So applied game theory concentrated on the one negotiation that mattered most and left the rest to experience.

A football transfer window shows the problem well. A club typically runs eight to twelve negotiations at once. A sale funds a signing, a signing releases a sale, and every completed deal tells the market what the club has left and how long it has to spend it. Inside the club, coach, recruitment and finance each want something different. The deadline does not move. Until recently, nobody would have tried to model that window as a single game: building the model would have taken months, and the window closes in weeks. An industrial buyer running a dozen coupled negotiations in one quarter faces the same structure with less drama.

AI removes that constraint. Specifying a game, keeping it current as facts change, and re-running it after every move used to be the expensive part. Now it is cheap enough to do continuously, across an entire negotiation period rather than for one workshop. A whole class of problems we used to leave to intuition becomes accessible.

Behavior may become easier to model

Game theory has always had an awkward relationship with reality. Its models are beautifully clean. Humans are not. People misunderstand incentives, care about fairness, retaliate, anchor, make mistakes, and refuse profitable deals because they are angry. That is why behavioral game theory exists in the first place.

As more strategic decisions are prepared by systems that evaluate alternatives, reactions and payoffs explicitly, at least parts of economic behavior may become more structured. Not perfectly rational. But easier to model. For decades we built increasingly sophisticated models and then asked why humans deviated from them. In the future, economic actors will increasingly use machines that were trained to reason strategically. The world might slowly move a little closer to the models.

---

Part two: what game theory does for AI

Here is the claim I would defend hardest, and it is stronger than “game theory is useful”.

If an AI is to give good strategic advice, it has to do game theory. There is no second option.*

Not because game theory is the best available method. Because it is the only language we have for the thing itself. A situation where my payoff depends on your choice, and your choice depends on what you expect me to do, has to be written down as players, options, information and payoffs before anything sensible can be said about it. That description is the game. A system that does not build it is not reasoning strategically, whatever it produces.

Everything else on offer is one of two things, and each fails at an identifiable point.

The first is optimization. Powerful, mature, and built on the assumption that the environment does not care what you do. It breaks the moment the other side reacts.

The second is prediction from data. Forecast what the counterparty will do, then respond to the forecast. This one is seductive, because it looks like it handles other actors. It doesn’t. It handles a record of other actors, and the whole point of strategic interaction is that their behavior is not independent of your move. A model trained on how suppliers behaved in past negotiations cannot tell you how this supplier behaves once it, too, has an AI anticipating you. Extrapolation fails exactly where the structure changes because of what you did.

So when a model is asked for negotiation advice today, it mostly retrieves what negotiation advice sounds like. That is not strategic reasoning. It is fluent, plausible and untethered from the other side’s incentives.

The research says the same thing in cleaner form. In a 2025 study in Nature Human Behaviour, Akata and colleagues let LLMs play repeated 2×2 games against each other, against scripted strategies and against humans. The models did well in self-interested games like the iterated Prisoner’s Dilemma. In coordination games such as the Battle of the Sexes they performed poorly, and they were strikingly unforgiving in repeated play. Performance drops further as strategic complexity increases. But once the models were given information about the opponent and prompted to reason explicitly about what that opponent would do, their behavior improved, including against human players.

The structure had to be supplied. It did not emerge.

You can call it agentic reasoning, simulation or scenario planning. If it works, it is game theory with the labels filed off. And if it does not represent the players, their incentives and their responses, it does not work — it only reads as though it might.

The harder problem is not the agent. It is the rules.

Calvano and co-authors showed in 2020 that simple reinforcement-learning pricing algorithms, competing in a standard oligopoly model, consistently learn to charge supracompetitive prices. Without communicating with each other, sustained by punishment strategies nobody taught them. No one programmed collusion. The rules of the game made it the rational outcome, and the algorithms found it.

Now scale that up: procurement agents, pricing agents, scheduling agents transacting at machine speed. What determines the outcome is not how clever any single agent is. It is the auction format, the information each side sees, the timing, and what a commitment actually binds.

Designing those rules so that strategically rational behavior produces a good outcome is mechanism design. It has always been the more powerful half of game theory, and it becomes more important as the agents get better, not less.

---

Part three: why both are great news for game theorists

Put the two directions together and the consequence for the discipline is not subtle.

Demand grows, because more decisions get recognized as what they always were. Supply grows too, because the constraint that kept applied game theory in a niche was never the theory. It was the cost of specifying and maintaining a model of a messy real situation. That cost is falling fast.

And the scarce skill moves. Solving a well-specified game is the part machines will do faster than any of us. Deciding which game is being solved is not, and it is where almost all the errors live. Who counts as a player. Which options are real and which are posturing. What each side actually knows. What a signed commitment binds and what it does not. None of that is in the data, because it is precisely the part that the data does not contain until someone has already made the decision.

I should be honest about the flipside. The work that used to be expensive — the modelling, the analysis, the calculation — is the work that gets cheap. If your value sits there, it is a hard decade. The value moves to specification and to design: getting the game right, and changing it when the one you are in produces bad outcomes for everyone rational enough to play it well.

Twenty-five years ago, applying game theory systematically to business negotiations was fairly unusual. Twenty-five years from now, strategic decisions may routinely involve humans and AI agents reasoning about other humans and AI agents. If that happens, understanding the game will not become less relevant. It may become one of the core skills behind intelligent decision-making.

---

* The obvious objection: poker and Go engines learn formidable strategic play through self-play, with nobody sitting down to model anything. Two answers.

Self-play needs the rules fully known and the situation repeatable millions of times. A supplier negotiation happens once, several of its rules are themselves negotiable, and the payoffs are private. That is not a shortage of compute. It is a different kind of problem.

And somebody did write the game down. The rules of Go are the specification, complete and available for free. Libratus and Pluribus were trained by counterfactual regret minimization — a game-theoretic algorithm — on a fully specified poker game. For two players that converges to an approximate Nash equilibrium. In the six-player case Brown and Sandholm dropped the equilibrium target and optimized for empirical performance instead, precisely because the solution concept stops carrying guarantees beyond two players. The game itself was still written down in full.

Nobody wrote the strategy. Somebody wrote the game. In a negotiation that step does not come free with the problem, and it is where the work is.

Sources

Akata, E., Schulz, L., Coda-Forno, J., Oh, S. J., Bethge, M. & Schulz, E. (2025). Playing repeated games with large language models. Nature Human Behaviour 9, 1380–1390.

Calvano, E., Calzolari, G., Denicolò, V. & Pastorello, S. (2020). Artificial Intelligence, Algorithmic Pricing, and Collusion. American Economic Review 110(10), 3267–3297.

Brown, N. & Sandholm, T. (2019). Superhuman AI for multiplayer poker. Science 365, 885–890.