Johns Hopkins is building classified versions of its AI wargaming tools for DoD, IC

LAUREL, Md. —  As the Trump Administration tries to jumpstart American AI amidst concerns about AI ethics and hallucinations, experts here believe they’ve found a low-risk, high-payoff way to leverage Large Language Models right now: wargames.

At least some federal agencies are receptive: Elements of the Defense Department, Energy Department, and intelligence community are asking the Johns Hopkins University Applied Physics Laboratory (JHUAPL) to take the AI-enhanced wargaming tools it’s developed, called GenWar and the Strategic AI Gaming Engine (SAGE), and upgrade them to run on classified networks using “highly classified data about our adversaries,” said James Miller, JHUAPL’s assistant director for policy and analysis.

“It’s going to be somewhat straightforward to move this onto TS/SCI networks,” Miller told a wargaming conference here last week, using the official shorthand for Top Secret/Sensitive Compartmented Information. “We have one sponsor that’s very keen to do it quickly and others are interested as well.”

A former undersecretary of defense himself, Miller preaches that “wargaming is a critical tool” for everything from long-term budget planning — for example, simulating how a proposed new weapon might affect a future war before investing billions to build it — to brainstorming the next move in the middle of a crisis. His fellow enthusiasts crammed into the conference to recount how they got senior Pentagon officials to throw handfuls of polyhedral dice to determine policy outcomes, or how they trained military officers in long-term planning by using poker chips to represent the command’s budget and then forcing them to improvise with random-events cards like “the Marine Corps grounds all V-22s.”

But such traditional wargames are labor intensive and time consuming, Miller and his fellow JHUAPL experts explained, which limits how often they can be used.

Counterintuitively, that’s also true of high-end simulations where the computer “plays” against itself, like the widely used Advanced Framework for Simulation, Integration, and Modeling (AFSIM). Yes, the model can run through multiple scenarios at machine speed — but only after humans have input all the proper parameters: the forces on each side, their tactics, the terrain, etc. Worse, that process is so complex and not user friendly that it often takes a team of specialists a month or more to set up a scenario.

“I took AFSIM training for a week and I still can’t do anything in it,” said Kelly Diaz, head of Advanced Concepts and Capabilities at JHUAPL. “It feels like you need a PhD.”

What about low-tech options? Many Pentagon wargames are just complex versions of the kind of board game on sale at hobby stores. Others resemble freeform roleplaying games like Dungeons & Dragons, where participants discuss their next move around a table and then ask a neutral adjudicator how it all turns out. While these games don’t require the technical skill of programming an AFSIM run, they can still take hours for the players and, behind the scenes, weeks of design and setup.

That’s time that Pentagon officials rarely have. So, Miller and his colleagues thought, what if they used AI to replace some, or all, of the human prep-time and players?  The two JHUAPL wargaming systems, GenWar and SAGE, apply this idea in two different ways.

GenWar works with the old-school high-fidelity models to make them easier to use. (Currently, that means AFSIM, but the plan is to add other DoD simulations in the future). In essence, GenWar replaces human scenario-builders with a customized chatbot. That allows a policymaker or commander to say what they want to simulate in plain English, and then a Large Language Model turns those words into inputs the simulation can understand in minutes, not months.

“It’s very intuitive,” said Diaz. “It’s a chatbot: You sit down and you talk to it.”

The chatbot portion of GenWar can hallucinate, of course. But the LLM isn’t running the simulation: It’s just acting as a translator between the human user and the actual high-fidelity model, which will reject any nonsensical inputs from the chatbot.

“It forces the AI to only do things that follow the laws of physics,” said Andrew Mara, head of national security analysis at JHUAPL. “It can’t spin off into ‘I landed 16 aircraft on the moon.’”

JHUAPL’s other AI-enhanced wargaming tool, SAGE, is in some ways even more ambitious: It uses generative AI to replace human players.

In it simplest form, a SAGE wargame simulates a National Security Council meeting or similar policy discussion, where various policymakers sit around a table and argue out a course of action — except that some, or all, of those policymakers are actually chatbots. SAGE can also simulate the staff advising a human player, the agencies executing their orders, the foreign countries responding to US actions, the adversary leaders opposing them, or even the neutral arbiter that compares both sides’ plans and adjudicates the outcome.

The immediate value of SAGE is as a training and brainstorming tool, allowing a single human to wargame on demand without assembling a whole team of players. But it’s also possible to have SAGE play itself, with every participant in the game being a chatbot.

Without human reality checks, the all-AI mode currently “loves to go off the rails,” Miller admitted to the conference. But even the crazier outcomes can provide the human user food for thought, he argued. And because the AI plays itself at superhuman speed, you can run hundreds of wargames in days, then look for patterns in the results.

“The goal isn’t to find ‘the answer,’” Mara told Breaking Defense. The value of AI wargaming is to explore a wider range of alternatives than humans could with assistance, discover strange outliers no human would have considered, and find recurring patterns, he said: “That’s where you point the humans and go, ‘hey … you really need to think through this.’”