AI Hallucination Nearly Triggered U.S. Boarding of Chinese Ship, Sources Say
A chatbot's false claim that a Chinese vessel carried nuclear components brought armed U.S. troops and aircraft to the brink of an interception before the error was caught.
The U.S. military came within minutes of boarding a Chinese cargo ship this spring after an AI chatbot incorrectly identified its freight as nuclear weapons components, according to an account first reported by The Decoder. Armed soldiers were staged and aircraft were already airborne when the mistake was discovered, halting the operation before it began. The near-miss has not been publicly confirmed by the Pentagon, and the specific unit, vessel and date remain undisclosed.
The sequence described is unusually concrete for a cautionary tale about AI error. A generated intelligence assessment, not a human source or sensor reading, appears to have driven the targeting decision far enough that troops were ready to put boots on a foreign-flagged vessel. That is the point at which a hallucination stops being a curiosity about chatbots and becomes an operational risk with diplomatic consequences. Boarding a Chinese ship on false pretenses could have escalated quickly, given existing tensions over shipping lanes and maritime claims.
The incident lands in the middle of a long-running argument inside the AI safety community. One camp worries about systems that become powerful enough to act against human interests on their own; another argues the nearer danger is ordinary, error-prone software placed in the hands of operators who do not check it, or who prefer not to. The reported episode fits the second camp squarely. No superintelligence was required, only a model that produced a confident wrong answer and a chain of command that treated it as actionable.
Military and intelligence agencies have moved quickly in recent years to fold machine-learning tools into targeting, surveillance and logistics, often with humans nominally in the loop. The friction is that a human reviewer under time pressure may rubber-stamp machine output rather than challenge it, particularly when the model presents its conclusion with the same fluency as a verified report. The reported near-boarding suggests the loop held only barely, and only at the last moment.
What happens next is largely opaque. It is unclear whether the episode prompted a formal review, a change to how AI-generated intelligence is labeled or corroborated, or any communication with Beijing. The open question is whether the near-miss produces stricter verification requirements before machine assessments can trigger military action, or whether it is quietly absorbed as a warning that arrived in time.