In spring 2026, a US intelligence report said a Chinese ship in the Middle East was carrying parts for a nuclear weapons programme. Armed personnel got ready to board it. Military planes were already in the air. Then, shortly before the operation, officials worked out that the report was wrong — and that a chatbot had helped write it.
CNN broke the story on 18 September 2026. Overnight, military AI intelligence reports went from a quiet policy question to a front-page argument. One source told CNN the mistake “almost started a war”.
So here is what a lot of people are now asking: does the military really let AI write intelligence reports? The short answer is yes, in part — and the safety net is thinner than most people assume.
Does the military let AI write intelligence reports?
Yes, in part. Pentagon staff can use approved AI chatbots to draft, summarise and check documents, including analysis work. The official position is that the AI assists a human, who stays responsible for the result. The September 2026 case shows an AI-assisted report can still travel a long way before anyone catches an error.
This is not a small pilot project. The Pentagon’s own AI portal, GenAI.mil, has passed 1.7 million users. The department has more than 3 million staff. So roughly half the organisation has a chatbot available at work.
What actually happened with the Chinese ship
According to CNN’s reporting, the sequence went like this:
- An intelligence report circulated across the US military in spring 2026, during the war with Iran.
- It claimed a Chinese vessel in the Middle East was carrying components linked to a nuclear weapons programme.
- The military prepared to intercept the ship. Armed personnel were ready to board. Aircraft were airborne.
- Just before the operation, officials learned the report had been produced with help from a chatbot, and that the AI had wrongly identified what the ship was carrying.
- A source described the report as “entirely false”. CNN could not establish what the ship was actually carrying.
Two details matter here. First, nobody hacked anything. No enemy planted the claim. The report was simply wrong, and it read like it was right. Second, it is still not publicly clear whether the analyst used a commercial chatbot or a government-built AI system.
The timing has made the story louder. Chinese President Xi Jinping is due in Washington on 24 September 2026, and AI is expected to be on the agenda.
Why did the AI get the ship’s cargo wrong?
The chatbot did not look the answer up. It predicted the words most likely to come next. When a model has no solid facts to work from, it still produces a smooth, confident sentence. Researchers call this a hallucination — an answer that is invented but reads exactly like a fact.
There is now a well-known explanation for why this keeps happening. In a September 2025 paper called Why Language Models Hallucinate, OpenAI researchers Adam Tauman Kalai, Ofir Nachum, Santosh Vempala and Edwin Zhang argued that the way AI models are tested rewards guessing.
Think of a multiple-choice exam with no penalty for a wrong answer. Guessing always beats leaving it blank. AI models are scored in much the same way. A model that says “I don’t know” earns nothing. A model that guesses is sometimes right and earns a point. So the training pushes it to answer every time.
That habit is the most dangerous thing about modern AI: it sounds exactly as certain when it is guessing as when it is right.
Which AI tools does the US military actually use?
Quite a few, and most of them arrived in the last two years.
- GenAI.mil — the Pentagon’s main AI portal. It launched on 9 December with Google Gemini. ChatGPT Mil and Grok for Government were added on 31 August 2026.
- ChatGPT Mil — cleared for “controlled unclassified information”, meaning sensitive government material that is not secret but still needs protection. It runs inside a government cloud, and officials have said they do not want these models used for combat.
- Ask Sage — a commercial platform cleared for work up to Secret and Top Secret level, running more than 150 different models behind one interface.
- CamoGPT — the Army’s own chatbot, first deployed in spring 2024, with around 75,000 users. It runs on the secret-level SIPRNet network and cannot browse the open web.
- NIPRGPT — the Air Force’s tool, which reached roughly 700,000 users for unclassified work. The Army blocked it in April over data-security concerns, and it was switched off on 31 December.
The money behind this is serious too. In summer 2025, the Pentagon awarded contracts worth up to $200 million each to OpenAI, Anthropic, Google and xAI.
The military did not stumble into AI by accident. It bought it, branded it and handed it to millions of staff.
Is a human supposed to check the AI’s work?
Yes. Every one of these tools is officially a helper. A human analyst is meant to verify what the AI produces before anything moves. The China ship case shows how that check can fail — the report travelled through the system until aircraft were in the air.
The reason is not simply that someone was careless. It is that AI writing is persuasive. A briefing produced by a chatbot uses the same calm, professional tone as a briefing produced by an expert who spent three weeks on it. There is no hesitation in it, no hedging, no “I think”. People trust the machine because it sounds sure of itself.
That same weakness turns up in ordinary office software. We covered a smaller version of it in Can Hidden Text in an Email Fool Your AI Summary? — when an AI reads something on your behalf, you inherit whatever mistake it makes.
Does any of this affect you?
In two ways, yes.
The first is direct. The same failure that nearly put boarding parties on a ship shows up in smaller, more personal decisions. An invented medicine interaction. A tax rule that does not exist. A court case that was never filed. A confident wrong answer about a visa deadline. The mechanism is identical. Only the stakes change.
The second is about rules. When an AI mistake nearly triggers a clash between two nuclear-armed countries, governments react. Laws written for military AI tend to spread outward into the tools ordinary people use. If you want a sense of how fast defence AI has moved, read our earlier piece on whether AI really flew an F-16 fighter jet by itself.
How to catch an AI answer that is made up
You do not need a security clearance to use the lesson. Five habits catch most invented answers:
- Ask for the source, then open it. If the link does not exist, or does not say what the AI claimed, you have your answer.
- Ask twice, in two fresh chats. If the two answers disagree on a number or a name, trust neither.
- Be hardest on specifics. Numbers, dates, names and quotes are where models invent most freely.
- Ask what it is unsure about. Many models will flag the shaky parts if you ask directly.
- Never let AI be the last step. Before anything expensive, legal, medical or permanent, a human checks it.
The Pentagon has the biggest AI budget in the world and still got caught out. A free chatbot on your phone is not going to do better.
Frequently Asked Questions
Did AI nearly start a war between the US and China?
According to CNN’s 18 September 2026 report, a false AI-assisted intelligence claim about a Chinese ship brought the US military close to boarding the vessel. One source said it almost started a war. No shots were fired, and the error was caught before the operation went ahead.
Which chatbot wrote the false intelligence report?
That has not been made public. CNN reported it is unclear whether the analyst used a commercial chatbot or a government-built AI platform. The Pentagon offers staff both kinds of tool.
Is the US military allowed to use ChatGPT?
Yes, through an approved version. ChatGPT Mil launched on the Pentagon’s GenAI.mil portal on 31 August 2026 for controlled unclassified work. It runs in a government cloud, and officials have said they do not intend these models to be used for combat.
Why do AI models make things up instead of saying “I don’t know”?
Because the tests used to score them reward guessing. A 2025 OpenAI research paper argued that saying “I don’t know” earns no points, while a guess is sometimes right. Across millions of training examples, that teaches the model to always produce an answer.
Can you tell an AI answer is wrong just by reading it?
Usually not. AI writes invented facts in the same confident tone it uses for correct ones. The only reliable check is to verify the specific claims — names, numbers, dates and sources — against a trustworthy outside source.
