This spring, during the war with Iran, an intelligence report went round the US military saying a Chinese ship in the Middle East was carrying components of a nuclear weapons program. Armed troops got ready to board it. Military planes were in the air. Just before the operation, officials looked again and found the report had been made with AI, and that the chatbot behind it had misidentified the cargo. CNN broke the story on September 18, based on four sources familiar with the episode.
We made a 46-second explainer about how a guess got that far. You can watch it on this page or as a YouTube Short.
What happened
According to CNN, an analyst at a special operations command asked a chatbot about intelligence on the ship's manifest that came from US Special Operations Command Pacific in Hawaii. The bot combined open-source information with secret signals intelligence and concluded the ship was carrying nuclear-program material. That was wrong. CNN could not learn what the cargo really was, or whether the chatbot was a commercial product or a government tool.
What we know, and what we don't
Then came the step that matters most. The analyst used AI a second time to package the answer into a standard intelligence report, the kind military officials trust, and sent it out. One source told CNN the report was "entirely false" and that it "almost started a war."
US Special Operations Command Pacific and the Pentagon did not respond to CNN's request for comment. Everything here rests on anonymous sources, so the film says "sources say" on screen and adds nothing CNN didn't report.
Why it got that far
The answer was wrong, but it didn't look wrong. It arrived in the format people already trust. That's the shift: making work look finished used to take effort, and the effort was a rough signal that someone had checked. That signal is gone.
OpenAI's GDPval test measures this on real work. It covers 44 occupations, and experienced professionals from those same occupations grade the AI's work against work by industry experts. In September 2025, Claude Opus 4.1 matched or beat the expert work on 47.6% of tasks, which the film rounds to 48%. In April 2026, OpenAI reported 84.9% for GPT-5.5, rounded to 85%. That second number is OpenAI's own claim about its own model.
The numbers behind the shortcut
Occupations in GDPval
Professionals from the same jobs grade the AI's work against work by industry experts.
Matched or beat experts
Claude Opus 4.1's wins or ties in OpenAI's GDPval paper. The film rounds it to 48%.
Seven months later
OpenAI's own figure for GPT-5.5 on the same test. The film rounds it to 85%.
The job now
I run AI on real finance work every day, and the pattern is the same at a desk as it is in a war room. Making the output look right is the easy part. The valuable part is the question before it goes out: is this true? And the earlier you ask it, the cheaper it is, because the next draft will look even better than this one.
Why a finance team should care
A board pack, a forecast or a variance commentary can now be produced in the house format in minutes. The format no longer tells you anyone checked the numbers. Before an AI-made document leaves your team, decide who checks the claims against the source, what counts as a source, and what happens when nobody did.
Sources. CNN, Katie Bo Lillis and Zachary Cohen, "US military had close call after using AI for false intelligence report, sources say", September 18, 2026. OpenAI, GDPval paper (September 2025) and Introducing GPT-5.5 (April 23, 2026). We checked all three live on October 7 and 8. The ship, the planes and the report in the film are drawn illustrations, not footage. The idea of a rising floor comes from Nick Saraev's video "Here's What I'd Learn Instead of AI Automation in 2027".
Three questions to ask
- Who checks the claims? Name the person who traces each figure back to its source, not just reads the draft.
- What counts as a source? A chatbot's answer is a lead to verify, never the evidence itself.
- What does the format promise? If your house template signals "checked", make sure someone did, before it goes out.
Looking finished is free now. Checking isn't.
The report travelled because it looked like every report the military trusts. That signal no longer means anyone checked, so ask "is this true?" early, before the next one looks even better.