AI hallucination nearly triggered US strike on Chinese ship, sources say

This spring, during the war with Iran, an intelligence report warning that a Chinese vessel in the Middle East was carrying nuclear weapons components pushed the US military to the brink of an armed interdiction. According to four sources, planes were airborne and troops were readying to board the ship before officials took a closer look and discovered the report had been assembled with help from an AI chatbot that misidentified the cargo. One source called the report ‘entirely false’ and said it ‘almost started a war’ — a confrontation with a nuclear-armed state that could have escalated fast. The analyst had fed the tool intelligence on the ship’s manifest, let it fuse open-source data with classified signals intelligence into a false conclusion, then used AI again to package the result into a standard, trusted report and disseminate it.

The near-miss lands as the Pentagon accelerates AI adoption across the force. Defense Secretary Pete Hegseth’s January ‘Artificial Intelligence Acceleration Strategy’ explicitly aims to strip away bureaucratic barriers and put commercial-grade models in the hands of three million personnel at every classification level. But the rollout is decentralized: different commands run different tools under different orders, with no unified standard for verifying AI-generated output and wide variation in reliability. A former senior official was blunt that the government’s internal systems are ‘mostly just copies of the commercial stuff wearing lipstick.’

The episode reframes the AI risk debate. While Silicon Valley voices warn about civilization-ending superintelligence, this case highlights a nearer and more concrete danger: humans acting on confident, fabricated machine output, especially as AI moves into targeting decisions where errors mean civilian casualties or fratricide. Sources say such hallucinations have not been isolated since these tools spread through the intelligence community, and that younger analysts — fluent with the tools — tend to trust them uncritically. As one source put it, ‘AI allows you to get to a bad idea faster.’