An AI error nearly caused a military incident between the United States and China

CNN could not determine what the ship was actually carrying.

An important clarification

The report spoke of components related to a “nuclear weapons program”. The available information does not allow us to say that the cargo was intended for the Iranian nuclear program. The incident occurred during the war with Iran, but these two elements should not be conflated.

How AI turned data into a false alert

The mechanism described by CNN shows that the error did not stop at a simple question posed to a chatbot.

An analyst from a Special Operations Command would have queried the AI using information about the ship’s cargo manifest. The initial intelligence came from the US Special Operations Command Pacific, based in Hawaii.

The system would then have combined open-source information with classified electro-magnetic intelligence held by U.S. authorities.

From this mix, the chatbot would have concluded, incorrectly, that the cargo included items related to a nuclear weapons program.

The analyst would then have used AI a second time to transform the result into an institutional standard intelligence report, before it was disseminated within the military apparatus.

This detail is central: an erroneous piece of information produced by a machine could take on the appearance of a document considered reliable by its recipients.

The used chatbot remains unknown

CNN could not determine which model produced the error.

It is not clear whether it was a commercial service, a tailored version of an existing product, or a tool developed specifically for the U.S. government.

An former senior American official familiar with the systems used by the military and intelligence services described certain internal tools to CNN as essentially copies of commercial products merely “re-skinned”.

This remark helps explain the difficulty of tracing the precise origin of a response generated in an ecosystem where commercial products, government versions, and internal systems coexist.

It does not, however, allow attributing the incident to Google, OpenAI, xAI, or another specific provider.

1.5 million people now use AI every day at the Pentagon

The case takes on another dimension when viewed in the context of how quickly artificial intelligence is spreading within the U.S. Department of Defense.

During a June 2026 appearance at the Hudson Institute, the Pentagon’s Chief Technology Officer, Emil Michael, said that 1.5 million people out of 3.5 million military personnel and employees were now using AI daily.

By December 2025, they were still only about 80,000.

So it is not merely about 1.5 million accounts created: Emil Michael explicitly spoke of daily usage.

From 80,000 to 1.5 million

In a matter of months, generative AI progressed from a tool used by a few tens of thousands to a technology integrated daily into the work of 1.5 million Pentagon personnel.

The department had launched its GenAI.mil platform in December to provide its personnel with controlled access to commercial AI models.

Google’s Gemini models were the first deployed, especially on the Pentagon’s non-classified networks.

Government versions of OpenAI’s ChatGPT and Grok then joined the platform at the end of August 2026.

This chronology is important: the Chinese ship incident dates from the spring. There is thus no basis to claim that ChatGPT or Grok were involved, nor that the tool used belonged to GenAI.mil.

A Congress report: 200 hours of work cut to five

Emil Michael himself gave in June a particularly revealing example of how these tools are used.

Some employees are required to produce each year mandatory reports for Congress. According to the Pentagon’s CTO, they can now upload the necessary documents into an AI tool and ask it to produce an initial draft.

A task that previously could require around 200 hours of human labor could thus be reduced to about five hours.

This productivity gain partly explains the Pentagon’s enthusiasm.

But the parallel with the Chinese ship incident is striking as well: in both cases, a machine absorbs a set of documents, synthesizes them, and then produces an institutional document intended for human decision-makers.

The difference obviously lies in the potential consequences of an error. A misstatement in an administrative report can be corrected; incorrect information in a military intelligence chain can contribute to triggering a real operation.

The Pentagon wants to go faster and faster

In January 2026, Defense Secretary Pete Hegseth had presented an “AI acceleration” strategy.

Its stated goal is to reduce administrative barriers, speed up experiments, and provide major U.S. models to millions of civilian and military personnel.

The Pentagon does not limit AI to administrative tasks.

In June, Emil Michael explained that all combatant commands now had access to AI capabilities and that uses could gradually evolve toward more complex functions.

He did note, however, that target lists continued to be developed independently of AI. The technology can help process large volumes of data to determine, for example, the best moment to strike a target already selected.

“One human in the loop”: is it a sufficient guarantee?

One of the frequently advanced arguments for governing military AI is to keep a human in the decision chain.

But CNN’s investigation precisely shows the potential limits of this principle.

A source familiar with Pentagon policies says that AI use in targeting is advancing rapidly and that there is not yet a sufficiently clear doctrine explaining how a human’s presence in the loop would concretely prevent civilian casualties or fratricide.

The Chinese ship incident illustrates the problem: humans were indeed present at every step, but they initially trusted a report whose central information had been generated erroneously.

It was ultimately other human officials, by deeper data verification before the operation, who identified the error.

The issue is therefore not merely: “Is a human present?”

The real question becomes: what does the human verify, with what data, according to what procedure, and before which decision step?

CNN mentions other “hallucinations”

The episode would not appear to be entirely isolated.

According to a CNN source, similar “hallucinations” have already appeared within the American intelligence community since the spread of these tools.

Several experienced officials interviewed by the network regard AI as also applying new pressure on analysts: since it allows faster production, it may encourage faster dissemination, sometimes before all traditional verifications have been carried out.

The younger analysts, more accustomed to AI tools, would also be prone to give them excessive trust, according to several CNN sources.

One summary of the risk by a source: AI can enable “arriving at a bad idea faster.”

Why this revelation comes at a particularly sensitive moment

The calendar adds a diplomatic dimension to the case.

CNN published its investigation on September 18, just six days before Chinese President Xi Jinping travels to Washington on September 24 to meet Donald Trump.

This visit is intended to cover, among other topics, trade, rare earths, Taiwan, the war with Iran, and the future of artificial intelligence.

According to Reuters, Washington and Beijing are due to discuss the possibility of establishing common AI safeguards, notably to limit certain risks of misuse.

U.S. Treasury Secretary Scott Bessent is also expected to discuss with Chinese Vice Premier He Lifeng the open and closed AI models before the Trump-Xi meeting.

The revelation of an incident in which an AI error nearly led U.S. forces to board a Chinese vessel thus provides a very concrete example of the kind of escalation risk the two powers must now manage.

An interception would have posed a major risk of escalation

It would be an overstatement to say that a war between the United States and China would necessarily have broken out if the operation had taken place.

This hypothesis remains impossible to prove.

One source interviewed by CNN nonetheless believes that the false report had “almost triggered a war”, given the risk that a U.S. military operation against a Chinese vessel would have posed.

The established fact is more precise: the preparations had progressed to the point where armed troops were ready to board and aircraft were already in the air before the operation was canceled.

The Pentagon and the US Special Operations Command Pacific did not respond to CNN’s requests for comment on this episode.

The new risk: giving an official-looking veneer to an error

The file ultimately highlights a different risk from the futuristic scenarios in which AI would autonomously decide to trigger a war.

Here, the machine did not make a military decision.

It produced an erroneous conclusion. An analyst picked it up. The AI then helped transform it into an institutional report. Officials received it and the army began preparing an operation.

The danger thus lies in a far more ordinary chain: a credible error, dressed in the format of an official document and transmitted to humans themselves under strong time pressure.

With 1.5 million personnel now using AI daily within the Pentagon, the question is no longer whether these tools will enter the U.S. military’s operating framework.

They are already there.

The challenge now is to determine which decisions can rely on them, which controls must be mandatory, and how far the pursuit of speed can go when an error could have immediate military or diplomatic consequences.