China's "97% Accuracy" Stealth-Killer Claim Quietly Admits It's Actually 90% on Mockups, Not Real Jets

 Researchers at the Beijing Institute of Technology, working with the China Airborne Missile Academy and a national key laboratory for near-surface detection, published a study describing a lightweight AI system designed to let heat-seeking air-to-air missiles recognize the infrared signatures of advanced fighter jets in milliseconds, distinguishing an aircraft's genuine heat signature from decoy flares released to break a missile's lock. The headline figure widely reported from the study, 97.1 percent recognition accuracy against mock-up F-22 and F-35 targets, comes from a training set of 3,245 infrared images captured by a missile-borne scanning system, with the model compressed to just 16.1 percent of the parameter count and 19.2 percent of the computational requirements of previous approaches, fitting onto a Zynq7020 hardware accelerator costing only a few hundred yuan and drawing just 2.2 watts, achieving inference in roughly 1.5 milliseconds.

That engineering achievement, genuinely fitting a functional AI recognition system into a power and size envelope small enough for missile integration, is real and worth taking seriously as legitimate technical progress. What's worth reading much more carefully, though, is what happens when EurAsian Times' own reporting pressed lead author An Jiangshan specifically on how that number applies to actual F-22s and F-35s rather than the training mockups. An's answer walked the headline figure back considerably: asked about real fifth-generation fighter identification, he said the relevant test data involved "confidentiality issues" and couldn't be disclosed, before adding that "for the simulated targets in the current test data set, the recognition rate can reach about 90 percent." That's a meaningfully different number than the widely circulated 97.1 percent, and the gap between the two, plus the fact that both figures apply to simulated mockups rather than any confirmed test against real F-22 or F-35 infrared signatures, is exactly the kind of caveat that tends to disappear between a research paper's careful hedging and the headline built from it.

This blog has now built a specific track record worth applying consistently here: earlier coverage of a Chinese MANPADS anti-stealth demonstration traced the drill's underlying premise back to an unverified Iranian F-35 claim the Pentagon never actually confirmed, and that same reporting cataloged a repeating pattern of Chinese anti-stealth announcements, bistatic space radar, the JY-27V meter-wave radar, BeiDou-based refraction detection, an IRST system claimed to track a B-2 from 150 kilometers, none independently verified and several directly contradicted by real-world performance. This new study's own source article supplies a particularly sharp data point in that pattern: the JY-27V radar this blog previously flagged as an unverified claim was actually tested in real combat conditions during the January 2026 U.S. raid that captured Venezuelan President Maduro, and it failed. Venezuela's entire layered air-defense network, JY-27V radars integrated alongside Russian S-300VM systems, collapsed as American aircraft operated freely through contested airspace, a direct, real-world refutation of exactly the kind of anti-stealth claim this new infrared study is now inviting similar confidence in.

The honest technical read here, distinct from the skepticism about verification, is that this specific approach addresses a genuinely real physics problem rather than a fabricated one. Stealth aircraft can reduce but not eliminate their heat signature, engine exhaust and aerodynamic friction heating remain physically present regardless of radar cross-section reduction, and a missile seeker sophisticated enough to distinguish that specific signature from decoy flares in under two milliseconds would represent a genuine, non-trivial threat to current countermeasure doctrine if it worked as described against real aircraft rather than mockups. The researchers themselves were notably candid about the study's limits, restricting their claims to close-range air-to-air fuzes, explicitly flagging that "the effect on surface-to-air missiles remains to be verified," and acknowledging more testing and data are needed, a level of hedging considerably more careful than the "97% accuracy against F-35s" framing that spread from it.

Reading this study against the JY-27V's actual combat failure and this blog's established pattern of treating unverified Chinese anti-stealth claims with real caution, the reasonable position isn't dismissing the underlying research, a functional, low-power AI infrared classifier is a genuine and useful engineering result, but it is refusing to extend that legitimate technical achievement into an unearned claim about operational capability against real F-22s and F-35s that even the study's own lead author, when pressed directly, wasn't willing to make.