A US military chatbot wrongly flagged a Chinese ship as nuclear cargo, nearly triggering a clash -- senators now want it investigated alongside the Minab strike

A US special-operations analyst this spring asked a chatbot to review intelligence on a Chinese cargo ship crossing the Middle East, then used AI a second time to turn the findings into a formal intelligence report, according to a CNN exclusive published Sept. 18. The chatbot had fused open-source information with classified signals intelligence and wrongly identified the ship's cargo as nuclear-weapons-program components -- and once the report circulated, it pushed the US military to the brink of an interception: armed service members prepared to board the vessel, and military aircraft were already airborne, before officials caught the error just before the operation began. One source described the report to CNN as "entirely false" and said it "almost started a war" with China.

Neither US Special Operations Command Pacific nor the Pentagon responded to CNN's request for comment on the incident. CNN could not determine what the ship was actually carrying, or confirm whether the chatbot involved was a commercial product or a government-built system -- only that it combined open-source and classified inputs into a single, uncorroborated conclusion that nearly reached a shooting war. Defense Secretary Pete Hegseth hasn't addressed this incident directly, but he told the Senate Armed Services Committee in an April 30 hearing that "we absolutely have to stay ahead" on military AI: "The advantage that AI provides applies to any number of capabilities, whether it's domain awareness, targeting cycles, you name it."

In a letter sent Sept. 19, Sens. Mark Warner, Jack Reed and Chris Coons told Hegseth and Director of National Intelligence Jay Clayton that recent events have given them "growing concern about the extent to which agencies under your oversight have prioritized acceleration of AI capability adoption and 'experimentation' over effective governance." The senators demanded "an immediate investigation by relevant Inspectors General, with a higher degree of public transparency in the ultimate findings," and asked that investigators get "unrestricted access to both instances this year in which media reports have suggested significant errors in AI-enabled targeting workflows" -- explicitly pairing the ship incident with the February strike on the school in Minab.

“AI allows you to get to a bad idea faster.” — a source with knowledge of the Pentagon's AI rollout, quoted by The Daily Beast

The senators' second cited instance is the more severe one. The Feb. 28 strike on the Minab school killed more than 150 people, according to a Wikipedia tally of the various investigations -- more than 120 of them children -- and a still-unreleased Pentagon internal review, first reported by Bloomberg and corroborated by IBTimes UK, found personnel at US Central Command "relied heavily on" Palantir's Maven Smart System, expecting the AI tool to flag outdated or inconsistent intelligence about the target. It didn't: an analyst had flagged changes at the site back in 2019, but that finding sat in a disconnected system and never reached strike planners, while more than 1,000 other Iranian targets were being struck in the same 24 hours, compressing the time available to double-check any one of them. The review also found that Hegseth-era staffing cuts had reduced the Pentagon's civilian-harm-mitigation teams by roughly 90% before the war began. Palantir has pushed back, saying it "was not responsible for the underlying data or for identifying intelligence deficiencies." (A separate UN fact-finding mission went further in September, concluding there were reasonable grounds the strike violated the laws of war -- a characterization the Pentagon's own review, as reported, does not use.)

What ties the two incidents together, beyond timing, is a gap in how the military verifies what its AI tools tell it. Analysts and former officials described a rollout that's decentralized by design -- different branches choosing their own tools under Hegseth's AI Acceleration Strategy, with no single verification standard across them -- and uneven in quality. "The internal tools are mostly just copies of the commercial stuff wearing lipstick," one former official told The Daily Beast. The strategy's own stated aim was to move fast: "eliminate bureaucratic barriers" and "unleash experimentation" across every mission area, including targeting and intelligence. Neither incident required the chatbot to be malicious or even unusual -- the ship report and the Minab strike both moved through processes that were, per their own reviews and reporting, working as designed, just without anyone positioned to catch a wrong answer before it reached people with weapons. This isn't the first time lawmakers have raised exactly this concern: at Hegseth's own April 30 hearing, Sens. Elissa Slotkin, Jacky Rosen and Kirsten Gillibrand pressed him for explicit human control over lethal-targeting decisions -- five months before either incident in this letter became public.

Both incidents are, on the record, ones where the system worked exactly as intended right up until the last check -- and in the Minab case, not even then. That is the senators' actual argument: not that the Pentagon's AI tools are unusually broken, but that "experimentation" at this speed, without a public verification standard, means the next uncaught error won't be a near-miss.