This spring, the US military came within one step of boarding a Chinese cargo ship in the Middle East. Armed personnel were already in position, aircraft were already in the air, and the operation was called off at the last stage. The reason: officials took a closer look at the intelligence report backing the interception and found it had been produced with the help of artificial intelligence — and the cargo it identified did not exist.
CNN broke the story on September 18, citing multiple people familiar with the matter. Other outlets followed, and the core facts held up: an AI-assisted intelligence product had traveled the entire chain from drafting to a live decision.
Two AI queries, one report
The report came from an analyst at US Special Operations Command Pacific, headquartered in Hawaii. The analyst first used a chatbot to check the cargo manifest of the ship in question; the tool blended open-source information with entries from classified intelligence databases to produce a conclusion. The analyst then used AI again to draft the formatted intelligence report. Stacked together, the two steps produced a document that reached the decision chain claiming the ship was carrying components bound for Iran's nuclear program.
The report circulated while the US and Iran were in active conflict. Under standard procedure, that kind of intelligence points to a maritime interdiction: send armed personnel aboard to inspect. Aircraft were already circling overhead. A military source told CNN the report was "entirely false, but almost started a war."
"Entirely false, but almost started a war."
The operation was never carried out. Officials ran a closer review before the plan was executed and caught both the report's AI origin and the tool's misreading of the cargo. CNN's reporting does not say what the ship was actually carrying, where it was headed, or which chatbot the analyst used. US Central Command did not respond to a request for comment.
Where the tool was plugged in
There was no jailbreak here, no adversarial attack, no hacker. The failure came from an ordinary use case: using a general-purpose chatbot to merge information. It stitched together open-source material and entries from classified databases into a conclusion that sounded complete and looked properly formatted — and that conclusion was then treated downstream as verified fact. The second step, AI-assisted drafting, added another layer of credibility: an intelligence product that read like something a person had written, in the right format.
The intelligence process normally includes multiple layers of human sign-off. The interdiction didn't happen because of an ad hoc closer look taken right before the operation, not because of a review checkpoint built into the process. Had that extra look not happened, or been done slightly more loosely, the document would have gone through as written.
Who bears the misjudgment
The target was a Chinese ship. Had the boarding actually happened, the consequences would have landed on the crew, the shipowner, and the two countries' crisis-management mechanisms — not on the analyst who wrote the report or the vendor whose tool was used. The technical risk sat inside a US process; the cost would have landed outside it. China has not publicly commented on the incident, and there's no reporting indicating Chinese authorities were notified.
For Chinese shipping companies running Middle East routes, the nature of the answer to "why did the US military flag my ship" is shifting: from a human-issued intelligence judgment to a document that may have passed through two rounds of AI. Which layer can be appealed, and who corrects an error, is not something any public mechanism currently covers.
Recommendations, for now, stay recommendations
The push to put AI into intelligence analysis, targeting decisions, and logistics has not been interrupted by this incident. Experts cited in public reporting have floated two ideas: drawing a line barring AI from nuclear-related systems, and setting up an AI-incident hotline between major powers, modeled on nuclear hotlines. Both remain at the proposal stage; neither has been written into any binding document.
Based on public information, nothing in the process has been formally changed yet. US Central Command has not said whether it has revised rules on analysts' use of general-purpose AI tools, or how many other documents beyond this one have gone through the same pipeline.
Sources: CNN Politics, Gizmodo, CocoLoop, Taipei Times. The command involved, the sequence of the two AI queries, and the "entirely false" quote were cross-checked against consistent reporting across multiple English-language outlets.