A chatbot misread a ship's cargo and troops got ready to board. Ask where a claim came from before it moves.

On September 18, CNN reported that a US military intelligence report, built with a chatbot’s help, wrongly said a Chinese ship was carrying parts of a nuclear weapons program, and that armed troops were getting ready to board it when officials found the error. The lesson is not “never use AI.” One cheap question, where did this claim come from and what else supports it, was asked at the end of the chain instead of the start.

This is an illustrative reading of CNN’s reporting through the Of One method. It adds no facts to the story and does not judge the people in it. CNN’s account rests on unnamed sources, and the Pentagon did not respond to CNN’s request for comment.

What was reported

According to CNN, citing four sources familiar with the episode:

The claim. This spring, during the war with Iran, an intelligence report circulated across the US military saying a Chinese ship in the Middle East was carrying components of a nuclear weapons program.
How it was made. A special operations command analyst asked a chatbot about reporting on the ship’s manifest. The chatbot fused open-source intelligence with secret signals intelligence and misidentified the material on board.
How it spread. The analyst used AI again to package the finding into a standard intelligence report, the kind military officials trust, and sent it out.
What it set in motion. The military planned an intercept. Two sources said armed personnel were preparing to board. Military planes were in the air.
How it ended. Just before the operation, officials dug into the report and found the error. One source called the report “entirely false.”

CNN could not learn what the cargo actually was. It was not clear whether the chatbot was a commercial product or a government tool. CNN also reported that there is no single standard for how the US verifies what these tools produce.

One source put the risk in a sentence: “AI allows you to get to a bad idea faster.”

The eight bones

Of One frames any decision with eight plain questions, the eight bones. Here they are filled in from the reporting, with what a decision map would have marked as solid, weak or missing.

Bone	Question	From the reporting	On a map
Decider	Who makes the decision?	Military officials planning the intercept (not named)	Implied
Goal	What outcome do they want?	Not stated; the plan was to intercept the ship	Implied
Facts	What is true right now?	A ship, reporting on its manifest, and a chatbot’s reading of that reporting	Weak: one claim, one machine inference
Limits	What are the hard limits?	CNN notes that any operation against a Chinese vessel could have risked armed conflict	Solid, and severe
Levers	What can they actually change?	An intercept and boarding; the reporting names no other option	Weak: one option
Causes	What will shift the outcome?	Whether the cargo is what the report says	Rests entirely on Facts
Proof	How will you know it worked?	Not in the reporting	Missing
Timing	When must they decide?	Not in the reporting	Missing

Look at the Facts row. The plan, the planes and the boarding party all stood on one claim about the cargo. On a map, every claim carries its evidence: where it came from, how fresh it is, how reliable. The honest entry here would have read “a chatbot’s conclusion from fused sources, not yet checked against the original reporting.” That is a guess treated as a fact.

The unknown nobody asked in time

The method gives every unknown a name, then asks the best next question: the one that could change the decision most, for its cost. Here it was cheap:

Does the original reporting on the manifest actually say this, and does any independent source agree?

Two details in CNN’s account put that question at the top of the list.

The chatbot fused its sources. It combined open-source intelligence with secret signals intelligence to reach its conclusion.
The format carried the trust. The finding went out in a standard report format that officials are used to relying on.

Of One requires a source independence pass for this reason. Two statements that trace back to one machine reading are one claim, not two, however many documents went into the reading.

The kill test

A kill test is the result that would prove you wrong. For this claim, a good one is quick and decisive:

If the manifest reporting the chatbot read does not name that material, the claim is dead until something else confirms it.

It needs no new collection, only a trip back to documents already in hand. By CNN’s account, officials did dig into the report, just before the planned operation. The check worked. It ran last, when the cost of being wrong was highest, instead of first, when it was lowest.

Where the human gate belonged

A human gate is a move a named person must approve. Of One puts one on moves that touch safety, public policy, physical safety, or that cannot be undone. This story has two places for one:

Before the report went out. Before an AI-derived conclusion enters a format others rely on, a named analyst signs that its key claim was checked against its source. The gate is about the claim, not the tool.
Before the boarding. Boarding a foreign vessel cannot be taken back. The gate there asks for one thing: the evidence behind the cargo claim and the result of its kill test.

Neither gate would slow sound intelligence by much. Both would move the check that eventually happened to the point where it was cheapest. The method does not promise that gates like these catch every error; it makes the missing check visible while there is still time to run it.

What to do this week

If you build or run AI tools that feed real decisions:

Tag every AI-derived claim with its source, and mark which claims are the model’s own inference rather than something a source said.
Write the kill test next to each claim that drives an action. If nobody can write one, the claim is not ready to act on.
Do not let a template carry trust. A polished report should show where each key claim came from, not hide it behind a familiar format.
Put a named person at every step that cannot be undone, and hand them the map, not the essay.

You can run the eight bones on a decision of your own with the free tool at of1.ai/ask.

If you want claim tags, kill tests and human gates built into an AI workflow your team already runs, that is the work of a consult: $400 an hour, in 5-hour blocks.