This spring, in the middle of the war with Iran, an intelligence report went around the US military: a Chinese ship in the Middle East was carrying components for a nuclear weapons programme. Planes were in the air, armed personnel were ready to board. Just before the operation someone looked closer. The report had been produced with a chatbot, and the chatbot had got the cargo wrong.
CNN broke the story on Friday with four sources. One of them calls the report «entirely false» and says it «almost started a war».
How a query became a finished assessment
An analyst at US Special Operations Command Pacific in Hawaii asked a chatbot about reporting on the ship’s manifest. The model fused open-source material with classified signals intelligence held by the government and reached its conclusion. The analyst then used AI again to pour the findings into the standard format of an intelligence report — the format officers trust — and sent it out.
Whether the tool was a commercial product or a government one, CNN could not establish. A former senior official offers a line that settles the question either way: the internal tools are mostly «just copies of the commercial stuff wearing lipstick».
Every corner of the apparatus uses something different
Defense Secretary Pete Hegseth issued an Artificial Intelligence Acceleration Strategy in January: cut bureaucracy, experiment, put AI in the hands of three million personnel at every classification level.
Execution is decentralised. Different parts of government run different tools under different orders and safety standards, and there is no single standard for verifying what those tools produce. One source says hallucinations like this one have not been isolated incidents across the intelligence community.
Targeting is where it gets ugly. «AI in targeting is definitely something that is ramping up», one source familiar with current policy says, «and there is no real guidance for how having a human in the loop will prevent civilian casualties or fratricide.»
The same day, Bloomberg reported that a Pentagon investigation into the February 28 missile strike on an Iranian school lists overreliance on AI among the failures. That strike killed 123 children.
The boring risk is the one that already happened
Washington has spent weeks arguing about models that break free of human control. This story is about something else: people treating a plausibly formatted falsehood as an assessment.
Anyone who works with these tools knows the pattern. A model answers in the shape you expect from a verified source, and the shape carries authority the content has not earned. On a normal day that costs you a wrong citation. Here it put planes in the air.
«AI allows you to get to a bad idea faster», one source says. Best summary of the year, and it didn’t come from a safety essay. It came from the Pentagon.