Over the past two years Chinese military researchers have moved from using simple replay analytics to training artificially intelligent agents that can replicate the decision patterns of human commanders and then place those agents in high-fidelity, multi-domain wargames. Beijing’s effort is not a gadget or a marketing stunt. It is an operational experiment at scale aimed at improving training realism, accelerating concept development, and stress testing doctrines shaped around intelligentized warfare.
Technically the work is straightforward but consequential. Researchers associated with the Joint Operations College at the National Defense University have published and presented methods that extract tactical and operational behaviors from recorded wargame replays and then train neural networks on that data to produce policy models. Convolutional neural networks and related architectures are used to learn maneuver patterns and force-employment tendencies from large replay datasets, producing agents that act with consistent doctrinal and cognitive biases. That approach improves adversary modeling fidelity because the agent reproduces human-like error, risk tolerance, and tempo preferences rather than behaving as an omniscient optimal planner.
Chinese publications and domestic reporting indicate the domain of application spans from unit-level tactical adjudication up to campaign-level “virtual commander” constructs that have been permitted extreme authority inside laboratory wargames in order to speed iterative learning. In public coverage the capability was described as being restricted to laboratory settings where the AI is tightly contained, a practical reflection of the Chinese principle that the Party and human commanders retain final authority over force employment. That containment is important politically, but it does not eliminate the operational value of granting AI wide latitude inside simulation. Researchers can run hundreds or thousands of permutations, harvest the resulting data, and refine both algorithms and doctrine in compressed cycles.
Western analysts and the U.S. Department of Defense have watched these developments closely and flagged them as part of a broader trajectory toward what the PLA calls intelligentized warfare. U.S. assessments have warned that Beijing’s operational concept emphasizes system-of-systems coupling, heavy use of simulations and exercises, and the development of subordinate operational concepts that rely on AI-enabled modeling to validate assumptions. The combination of doctrinal emphasis and dedicated research facilities suggests the PLA intends to institutionalize AI-augmented wargaming as a core element of future operational planning.
There are three operational advantages the PLA is chasing through AI-enabled war games. First, scalable adversary replication. AI agents trained on replay data allow a single simulation cell to emulate multiple culturally and doctrinally distinct adversaries, improving Blue-Red friction modeling without a large cadre of specialist umpires. Second, cycle-time compression. Algorithms adjudicate and evaluate many more iterations per calendar week than human-only teams can accomplish, accelerating concept maturation and tactics refinement. Third, measurement and repeatability. AI-driven runs produce structured telemetry at scale which makes statistical analysis of outcomes feasible, moving wargaming closer to data-driven experimentation. These are not hypothetical benefits. They are the explicit motivations in open Chinese technical literature and visible in the architecture of PLA wargame centers.
Those advantages are counterbalanced by three acute risks that both Chinese authors and outside observers have raised. The first is overfitting to historical play and simulation artifacts. An agent trained on replay data will replicate the biases and mistakes present in that dataset and may be brittle when confronted with novel cross-domain surprises. The second is escalation blind spots. Simulation runs that grant AI agents greater authority than would exist in reality can produce courses of action that appear plausible inside the model but that would be unacceptable or escalatory in real operations. Several independent wargame studies in allied circles have shown AI or automated adjudication can produce more aggressive choices under certain assumptions. That pattern suggests an important danger when AI-derived tactics inform real-world posture or crisis signaling.
The third risk is institutional. If militaries lean on AI adjudication to the point that human judgement atrophies, the value of wargaming as a training tool diminishes. Several critiques of AI-augmented wargaming argue that overreliance on algorithmic decision support can de-skill officers and obscure causal relations in favor of opaque correlational outputs. The practical implication is that AI should be used to expand exploratory breadth while preserving human-led adjudication and critical review.
For Western planners these Chinese developments are a call to recalibrate, not to panic. Three concrete steps should guide policy and procurement in response. First, invest in red-team simulation capability that can emulate Chinese AI-augmented agents. If Beijing is training virtual commanders, NATO planners must be able to subject their concepts to the same class of adversary models so decision cycles and escalation dynamics can be properly stress tested. Second, require transparency and auditability in any AI adjudication tools used for operational planning. Models must expose their training data lineage, key failure modes, and sensitivity to scenario framing. Third, preserve and emphasize human-in-the-loop training. Use AI to augment adjudication throughput and to highlight edge cases, but keep final operational judgment and crisis communication squarely in human hands.
Finally, the PLA’s use of AI in wargaming is a strategic signal. It demonstrates how doctrine, training, and technical R&D can be fused through military-civil integration to accelerate capability cycles. The United States and its allies should treat Chinese AI war games as both a capability to be measured and a methodology to be countered. That means better measurement of Chinese simulation outputs, more rigorous cross-domain red teaming, and shared allied standards for responsible use of AI in operational exercises. The murky part of the future is not whether AI will be used in wargames. It is how militaries will manage the transition from improved simulation fidelity to real world decision making without creating new incentives for miscalculation.