China showcases AI planner for large air-strike exercises

The reported gains are developer-reported test claims

Deng’s account says the first full-process test produced failures and missed performance requirements with 15 days remaining before an exercise. He said the revised system’s final test completed planning three times faster than the requirement and achieved coordination accuracy 20% above its target; he also said scenarios that once took days and nights of calculation could then be completed in minutes. The same account says that, in a 2022 high-altitude test, more than 100 tactical units coordinated strikes with the system.

Those numbers are meaningful only within the developer’s own reported test setting. The public accounts do not disclose the requirements, accuracy metric, comparison method, system inputs or an independent evaluator. They therefore support a claim that the team reported meeting its targets, not a like-for-like measure of operational effectiveness.

The source record also supplies an important limit on any claim of maturity. A separate profile of the unit says its software is refined through repeated matching tests and upgrades before it goes to troops. Its description of an air-refuelling assistance system, a large-scale strike system and airport-pavement monitoring software shows a broader effort to put software through that path; it does not independently validate the disclosed planner’s results.

A unit shaped by test-and-evaluation work

The unit’s history helps explain why its claims are worth watching but not treating as proof of deployment. Its profile traces its lineage to a force that built China’s first comprehensive missile test range in the Gobi Desert in the 1950s. It says the J-10 and J-20 were tested there before joining service.

That is relevant test-and-evaluation pedigree. It does not identify the planner’s architecture, the data it uses, the degree of human override or how it behaves when communications and assumptions fail.

The same profile says pilots train against AI air-combat agents in a virtual-combat system and that the unit is exploring human-unmanned teaming and AI-assisted decision-making. These are adjacent lines of training and experimentation, not evidence that the strike-planning system itself performs those functions.

What would show operational value

The next evidence needed is less a new headline than a clearer boundary around the tool: which decisions it recommends, which remain with commanders, what data and constraints govern its plans, and how its results compare with a stated baseline.

Until that is public, the defensible conclusion is limited. China is using AI planning software in increasingly complex air-force exercises and its developer reports substantial test gains. The available reporting does not establish independent reliability, combat use or autonomous control of aircraft or weapons.