In a structured wargame designed to test tactical decision-making under pressure, AI agents powered by large language models maintained more consistent strategies than experienced human officers, who drifted from their own plans as fatigue and information overload set in. That finding, drawn from a pair of academic studies published in early 2025, represents some of the most detailed public evidence to date that AI systems can sustain battlefield-level coherence in scenarios where human teams falter.
The research lands at a charged moment. China’s People’s Liberation Army has spent the past two years publicly signaling its intent to weave artificial intelligence into command and planning structures, and the studies, while not produced by or for the PLA, describe exactly the kind of capability Beijing says it wants. Whether lab-grade performance translates to real combat remains an open and critical question, but the direction of the technology is no longer speculative.
What the studies actually show
The first study, “Human vs. Machine: Behavioral Differences Between Expert Humans and Language Models in Wargame Simulations,” hosted on the arXiv preprint platform, pitted LLM-driven agents against teams of expert human players across multiple rounds of a structured wargame. Researchers measured behavioral consistency, tracking how reliably each side executed its own declared strategy over time. The AI agents showed markedly less variance. Human teams, by contrast, made increasingly erratic choices as rounds progressed, a pattern the authors attribute to cognitive fatigue and the compounding difficulty of processing battlefield information in real time.
The second, “Command-agent: Reconstructing warfare simulation and command decision-making using large language models,” published in the Elsevier-indexed journal Defence Technology, goes further. It describes a framework that pairs LLMs with digital-twin environments to run assault-level command simulations. Within that system, an AI agent issues orders, adapts to shifting conditions, and coordinates virtual units at speeds no human staff section can match. The paper presents a general architecture, not a classified PLA program. Its technical approach, however, resembles the AI-enabled command concepts that PLA-affiliated publications have described as strategic priorities, an alignment that is suggestive but interpretive, since the paper itself does not claim a direct PLA connection.
A third source adds institutional context. The U.S. National Defense University Press published an analysis surveying PLA Daily writings on AI’s military potential. That review catalogs both the ambition and the anxiety visible in official Chinese military media: enthusiasm for AI-driven speed and precision alongside stated concerns about training-data quality, vulnerability to deception and camouflage, and the risk that machine-speed decisions could trigger unintended escalation.
“The consistency gap between AI and human teams was one of the clearest results,” said a co-author of the arXiv wargame study in a written response to questions about the findings. “But consistency is not the same as correctness. A system that reliably executes a flawed plan is not superior to a human who recognizes the plan is wrong.”
What remains unverified
No declassified record from PLA operational exercises has surfaced, as of May 2026, confirming that a Chinese military AI system has outperformed human commanders in a live or hybrid assault scenario. The academic papers test general-purpose LLMs in abstracted environments, not PLA-tailored tools running against real terrain data, electronic warfare conditions, or adversarial countermeasures.
Direct, attributable statements from Chinese military officials on how far AI integration has actually progressed are absent from the public record. The NDU analysis draws on PLA Daily editorials, which reflect institutional messaging rather than operational reporting. Those editorials flag problems with no easy fix: training data that may not reflect real battlefield conditions, AI systems that can be fooled by deliberate misinformation, and escalation dynamics that become harder to manage when machines compress decision timelines from hours to seconds.
U.S. intelligence assessments of PLA AI trials, if they exist, have not been declassified or leaked in a form independent researchers can verify. News coverage framing AI as already “topping” human commanders in Chinese military exercises typically relies on secondary interpretation of the same academic and media sources described here, not on independent battlefield evidence.
How to weigh the evidence
The two academic papers are primary evidence of AI performance in simulation. Both use repeatable experimental designs, disclose their methods, and are available for scrutiny. The arXiv preprint has not completed formal peer review; it is hosted on a repository supported by major university members and carries the transparency expected of open-access research. The Defence Technology paper sits in a journal that applies Elsevier’s editorial and review standards.
The NDU Press analysis occupies a different tier. It does not generate new performance data; it synthesizes what PLA-affiliated writers have said publicly and evaluates those claims against known technical constraints. Its value lies in documenting the internal skepticism that headline-level coverage often omits.
Readers should treat the simulation results as proof of concept, not proof of battlefield superiority. An LLM that holds a coherent strategy in a bounded wargame has demonstrated a real and measurable capability. But that wargame operated with clearly defined rules, limited unit types, and transparent scoring. Real operations involve incomplete information, conflicting objectives, and political constraints no current simulation fully captures. Human commanders must weigh civilian harm, alliance politics, and long-term deterrence effects, none of which are easy to encode into a reward function or scenario script.
The studies themselves highlight trade-offs rather than a clean AI victory. LLM agents excel at consistency and rapid adaptation but depend heavily on the quality of their inputs. Poorly specified objectives or misleading scenario descriptions can push them toward brittle or escalatory choices. Human teams may be slower and less consistent, but they can sometimes recognize when a scenario itself is flawed or when a formally “optimal” move is politically or ethically unacceptable.
The broader competition
China is not working in isolation. The United States has invested heavily in its own AI-enabled command initiatives, including the Combined Joint All-Domain Command and Control (CJADC2) framework and DARPA’s Air Combat Evolution (ACE) program, which demonstrated AI-piloted fighter maneuvers in live flight tests. NATO allies, including the United Kingdom and France, have published their own AI defense strategies. The competitive dynamic matters because it shapes how quickly each side feels pressure to move AI from advisory roles into more autonomous decision-making loops.
The practical takeaway for defense planners and the broader public is that AI-enabled command is no longer theoretical. Working prototypes exist in academic settings, and at least one major military power is openly discussing how to operationalize them. The risk is not that AI will replace human generals overnight but that the race to deploy these systems could outpace the development of safeguards against their known weaknesses, including data-quality failures, susceptibility to spoofing, and compressed escalation windows.
In the near term, the most likely outcome is a hybrid model in which AI systems serve as fast-running advisers rather than autonomous commanders, generating options and forecasts that human officers must still interpret and approve. How effectively militaries manage that division of labor, deciding when to trust machine recommendations and when to override them, may matter more than raw performance scores in any single wargame. The research record as of spring 2026 shows that AI can compete with human experts in narrow simulations. It does not yet show that machines can bear the full weight of real command responsibility.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
