Should I use AI in my wargame?

But while AI remains in the “glue on pizza” stage of its evolution, and perhaps for some time afterward, there will be risks for the wargamers who use it. These risks will vary in scope and severity, and game producers and consumers must understand how and why.  

One of the first principles of gaming is that “you can game a lot of things, but you need to be specific about what question you are exploring.” This simple piece of wisdom has stood us in good stead through more than 20 years (combined) of gaming experience.

With all games, we begin by asking whether the question covers a topic that is fundamentally “novel.” That is, has the topic been studied in some depth, such as operational warfighting or nuclear escalation, or does it remain poorly understood, like integrated deterrence? We then ask about the type of game needed to explore the question: can we tweak a gaming method we already have, or will we need something new?

This allows for a neat two-by-two matrix (which, as Ph.Ds in the social sciences, we believe we are morally, if not legally, obligated to use wherever possible):

Those who see AI as a useful tool often highlight its ability to quickly run many versions of a game. AI can explore variations on established results and produce data to be analyzed using non-gaming methods. Done with well-understood topics and game systems, this use falls into the top-left box.

Those who argue that AI can help brainstorm new mechanics, new ways to game a topic, or create materials for a game on the other hand, are generally talking about building a new tool to look at a problem. Done for familiar topics, this falls into the bottom-left box.

We believe that these are both relatively sound applications of current and near-future large language models and other forms of “artificial intelligence”—because we can cross-check results.

Risks become greater when you explore unfamiliar questions, whether you use AI or not. Layering an emerging, fast-changing technology onto a poorly understood topic, even when you believe you know the game system—the top-right case—raises a host of risks. The dangers are even greater when you are building a totally new system to explore a poorly understood problem set—the bottom-right box.

A second layer of complexity in introducing AI into a wargame turns on whether you are mostly interested in quantitative data—that is, the results of the game and the specific actions players made (or did not make)—or in qualitative knowledge: the interaction and data sharing between players during the game.

When we glean quantitative data from a game about a known topic, the AI tools of today and the near-future can be useful in post-game analysis. The problem is already well bounded, and we have many methods with which to assess the results.

But if we are seeking to understand why people make certain choices, AI-infused games are potentially misleading. Consider the discussions around AI being used as a player or to serve as an “adversary-in-a-box.” Humans are naturally good at reading other humans but often struggle to correctly interpret actions from machines, making AI-driven interactions tricky. If you replace actors with AI in a game, you risk learning more about human-machine interaction than anything about the game or problem.

Formal red-teaming methods remind us there are actually at least four “entities” in any given interaction: Person X, Person Y, and each’s mental model of the other. As anyone who has misunderstood another’s perspective can relate to, problems tend to occur when one person’s mental model of the other diverges from reality.

Intuitively, we understand that a human-emulation cell, no matter how expert, is not the same as the real actor. When moving from a human assessment of the self and the other to one of a machine, we add in an additional layer of divergence and we risk false precision in trusting simulated actions. Much as maps distort reality by playing with our perceptions, we risk the same bias with AI.

From the authors’ perspective, much can be learned from observing both the outcomes of games and the interactions from players. Yet we learn not just from game results or discussions, but also from interactions between players and game designers. For this form of knowledge-generation, which is perhaps entirely experiential, the ways in which we can use AI to generate knowledge is even more unclear. What do we learn about the problem at hand? What do we learn about the game itself? What do we learn about wargaming as an enterprise? If the purpose of wargaming is to generate learning, we must interrogate what can be learned from this interaction between the human designer, the AI player, and the game itself.

Wargames and novel technologies have a troubled history. Still, in settings where we have clear measures of effectiveness based on other tools and can cross-check results, AI can help the wargaming community expand its reach. If, however, AI is used to replace human judgement—for example, to “impersonate” an adversary decision-maker or to replace human interaction—we risk a dangerous mirage of knowledge, and there be dragons.

Stephen M. Worman is director of the RAND Center for Gaming, a political scientist at RAND, and a professor at the Pardee RAND Graduate School.

Bryan Rooney is a political scientist at RAND.