Chatbot-generated intelligence report nearly triggered military confrontation with China

A hallucination in an AI-assisted intelligence report sent the US military scrambling to intercept a Chinese vessel in the Middle East last spring, convinced it was transporting components for a nuclear weapons program. Four sources familiar with the episode say operational plans advanced to the point where armed troops were preparing to board and aircraft were already airborne. Only at the last minute did it emerge that the report, circulated by Special Operations Command, had been produced with a chatbot that misidentified the cargo. The assessment was "completely false," one source said, but "almost started a war."

The analyst at Special Operations Command Pacific, based in Hawaii, had queried the chatbot about intelligence reporting on the ship's manifest. It is unclear whether the tool was a commercial product or a government system; a former senior official described internal tools as "copies of the commercial stuff with lipstick." The bot fused open-source intelligence with classified signals intelligence (SIGINT) from government databases and drew the wrong conclusion about the cargo. The analyst then used AI a second time to package the findings into the standard intelligence-report format that officers trust, and distributed it up the chain.

The episode exposes the risk of rushing large language models (LLM) into operational decision-making. In January Defense Secretary Pete Hegseth released the Pentagon's "AI Acceleration Strategy," which mandates making the technology available across every service — from target selection to logistics and budgeting. The stated rationale: the military must decide faster than potential adversaries such as China. In practice, the strategy pushes broad deployment before verification and explainability mechanisms have matured.

The problem is not only a one-off hallucination. When an analyst uses a chatbot to interpret raw material and then a second tool to draft the final product, the chain of trust breaks twice: the first model may invent connections that do not exist in the data, and the second may wrap the error in a professional shell that obscures its origin. Special Operations Command and the Pentagon declined to comment to CNN. Meanwhile, the intelligence community continues to embed the technology in every corner, without a clear kill switch for when a model errs with full confidence.