Can large language models reliably assist military strategists in planning warfare scenarios and testing defensive strategies? The U.S. Department of Defense is betting yes. Johns Hopkins Applied Physics Laboratory (APL) is launching the GenWar Lab in 2026, a facility designed to accelerate military wargaming by embedding generative AI into tabletop exercises traditionally conducted by human planners and strategists. While the promise is compelling—faster scenario iteration, AI-assisted decision analysis, and 70% to 80% solutions that enhance human learning—the stakes are extraordinarily high. Strategic failures in wargames directly influence national security policy, force posture, and wartime decision-making. This creates a critical tension: LLMs excel at pattern matching and natural language, but they operate as statistical models without ground truth validation, reliable reasoning chains, or proven benchmarking against real military strategy and statecraft. The GenWar Lab initiative raises essential questions about AI trustworthiness in life-or-death military contexts and the risks of reducing strategic analysis to “what the AI said.”
The GenWar Lab marks a significant departure from traditional wargaming methodology. Since the Prussian Army’s kriegsspiel in the 19th century, tabletop exercises have relied on Blue versus Red teams with human umpires adjudicating outcomes. This approach, while proven, is labor-intensive, difficult to iterate, and slow to incorporate lessons learned. The GenWar Lab introduces three new architectural components: GenWar TTX (Tactical Training eXercise), which creates the digital environment and AI agents; GenWar Sim, built on the government’s Advanced Framework for Simulation, Integration and Modeling (AFSIM), which allows human players to interact directly with physics-based models; and a translation layer powered by large language models that converts human speech into mathematical commands and vice versa. This enables humans to say, “I’d like to attack here, I’d like to defend there,” and have those commands automatically executed by sophisticated modeling engines without requiring operators to understand underlying equations. The lab explicitly contemplates wargames conducted entirely by AI actors on both sides, with no human participants. Kevin Mather, who heads the GenWar Lab, stated the facility will open in 2026 at Johns Hopkins APL in Laurel, Maryland. The lab responds to what Mather calls a “demand signal” from military sponsors wanting faster, deeper wargaming with more iterations and what-if analysis capability. By allowing scenario replay, decision logging, and post-game analytics on a digital foundation, GenWar aims to reduce the friction that has historically slowed strategic learning.
The stakes for GenWar are unusually high because wargaming directly informs operational doctrine, force design, and strategic policy. Military leaders use wargame outcomes to justify budget priorities, force structure decisions, and strategic pivots. Flawed conclusions from poorly designed or poorly validated AI-assisted exercises could cascade into real wartime errors—misjudgments about an adversary’s behavior, overestimation of own-force capabilities, or failure to anticipate second- and third-order effects. The lab’s promise is compelling: faster iteration means more scenarios can be tested, more strategies can be explored, and human strategists can accelerate learning through AI-assisted analysis. But generative AI, by design, does not guarantee correct answers. LLMs are statistical models trained on patterns in text, not on verified military science or validated strategic principles. They can hallucinate facts, conflate tactics, and produce seemingly plausible but strategically nonsensical advice. Benjamin Jensen, a researcher at the Center for International and Strategic Studies, articulated the core risk: “The larger challenge is that most foundation models commonly used haven’t been sufficiently benchmarked against strategy and statecraft.” If AI becomes the primary decision-maker in wargaming—if scenarios are run entirely by AI actors without human scrutiny—the exercise risks becoming an echo chamber of LLM assumptions, not a genuine test of strategy. The risk that military planners will accept an AI-generated recommendation without rigorous validation, simply because it came from a sophisticated system, is non-trivial. Additionally, LLMs have no accountability mechanism and no audit trail that senior military leadership can trace when evaluating strategic recommendations. The human learning curve in wargaming depends on transparent, documented reasoning—exactly what LLMs often lack. Finally, there is the institutional risk: over-reliance on fast, AI-generated wargame conclusions could erode the kind of deep strategic analysis that has historically distinguished military planning from commercial business strategy.
The GenWar Lab architecture combines three operational layers. The first is scenario definition and AI agent generation. GenWar TTX uses LLMs to automatically generate AI agents that play roles in the exercise: staff advisers, enemy commanders, civilian officials, or allied leaders. These agents can be given briefing documents, historical precedents, or strategic objectives, and the LLM generates realistic dialogue and decision-making patterns. A human player can interact with these agents in plain language, reducing the overhead of coordinating complex multi-player exercises. The second layer is command translation and model execution. When a human player states a tactical intent—”defend the valley with mobile reserves”—the GenWar Sim layer must translate that into mathematical inputs for the underlying physics model (AFSIM). This is where the true engineering innovation lies. Rather than requiring military operators to understand differential equations or modeling frameworks, the LLM acts as a bridge, parsing human intent and executing the corresponding commands in the simulation engine. The system maintains a digital log of all decisions, allowing players to rewind, retry, and replay scenarios. The third layer is adjudication and outcome analysis. Instead of a human umpire ruling whether an attack succeeds, the physics model calculates outcome probability based on factors like terrain, supply lines, air support, and force composition. An LLM then translates those numerical results back into plain language: “Your defensive position was breached due to insufficient air support and lack of second-line fortifications. Enemy casualty rate: 40%. Your unit integrity: 65%.” This feedback loop allows rapid iteration. A strategy that fails can be replayed immediately with a different approach, with the system capturing and comparing outcomes. The lab explicitly envisions scenarios in which both Blue and Red are controlled entirely by AI actors. In this mode, the human role shifts to observation and post-hoc analysis: watching the AI-conducted exercise, identifying unexpected outcomes, and asking “why did that happen?” The hope is that this mode will allow faster scenario exploration and the ability to test large numbers of strategies without human fatigue or attention constraints. However, this mode introduces the validation problem: without humans in the loop making real-time judgments about AI behavior, how do strategists know whether the AI exercise is producing strategically relevant results or merely internally consistent fiction?
Wargaming has been central to Western military thinking since the 19th century. The Prussian General Staff used kriegsspiel to train officers in operational reasoning and to test offensive doctrines. After World War II, the U.S. military institutionalized wargaming as a tool for strategy validation, nuclear deterrence analysis, and force structure planning. During the Cold War, wargames tested NATO’s ability to defend against Soviet invasion, evaluated the credibility of nuclear deterrence, and explored scenarios that had never occurred and might never occur. The rigor of traditional wargaming lies in the presence of skilled umpires who understand military history, doctrine, and the constraints of real warfare. A good umpire can say, “That tactic won’t work because you’ve neglected supply lines,” or “Your unit wouldn’t behave that way under artillery fire.” This institutional knowledge acts as a check on unrealistic player assumptions. The shift toward AI-assisted wargaming reflects broader Pentagon trends: digitizing decision-making, accelerating planning cycles, and reducing dependence on human expertise that takes years to develop. In an era where adversaries (particularly China and Russia) are investing heavily in AI for command and control, there is institutional pressure within the U.S. military to “move faster” and not be left behind by AI-driven competitors. However, faster is not always better in strategy. The fiasco of Iraq war planning—where strategic assumptions were validated by echo-chamber analysis and groupthink—demonstrates the risk of moving quickly without rigorous adversary analysis and red-team skepticism. GenWar Lab designers are aware of this. Program manager Kelly Diaz emphasized that the system is meant to “accelerate human learning,” not replace human strategists. Mather stated explicitly: “It’s not nearly as in depth, for example, as a traditional ops analysis or modeling and sim study. But AI can offer those 70% to 80% solutions that are not the answer, but really accelerate the human learning.” This framing suggests GenWar is intended as a brainstorming tool, not a strategic oracle. But the history of technology adoption in the Pentagon suggests there is risk of mission creep: a tool intended to support human decision-making gradually becomes the primary decision-maker. The absence of clear institutional boundaries and validation standards for AI-generated wargame outcomes creates this risk.
Primary Source: Defense News: “New Lab Offers Generative AI for Defense Wargaming” (Nov 24, 2025) — Full article detailing GenWar Lab architecture, LLM integration, AFSIM framework, and quotes from Kevin Mather (GenWar Lab Director) and Kelly Diaz (Program Manager).
Johns Hopkins APL Official Announcement: Johns Hopkins Applied Physics Laboratory: “GenWar Lab” (official news release) — Formal announcement of GenWar Lab opening in 2026, facility location in Laurel, Maryland, and research objectives.
Critical Analysis Source: Benjamin Jensen (Center for International and Strategic Studies) — Direct quote in Defense News article emphasizing the need to benchmark LLMs against military strategy and statecraft, and the risk of reducing strategic analysis to “Here is what an LLM said.”
Historical Context – Kriegsspiel: Center for Army Analysis reference to Prussian kriegsspiel methodology (1812 onwards) — Traditional approach to wargaming that serves as baseline for modern military exercises.
Key Technical References: Advanced Framework for Simulation, Integration and Modeling (AFSIM) — government-owned modeling and simulation framework underlying GenWar Sim.
Operational Security and Validation Concerns: The absence of published validation standards for AI-generated wargame outcomes in DoD guidance documents creates institutional risk. The GenWar Lab team has stated that exercises “will not remotely claim that [AI actors] are making optimal decisions,” yet the precise definition of “realistic enough” for strategic decision-making remains undefined.
Recommended Defensive Controls: (1) Explicit institutional mandate requiring all AI-generated wargame recommendations to be independently validated by subject matter experts with mandatory dissent mechanism; (2) Complete audit trail and decision documentation for all AI commands and outcomes, retained for 10+ years for post-conflict analysis; (3) Red team protocol: dedicated human team tasked with identifying implausible or strategically dangerous AI recommendations before they influence policy; (4) Benchmarking protocol: regular testing of GenWar outputs against known historical military scenarios to validate predictive accuracy; (5) Sunset clause: automatic re-evaluation every 3 years of whether AI-assisted wargaming has actually improved strategic decision quality compared to traditional methods.