US Military Nearly Attacked Chinese Ship Over AI Misinformation, Report Says

It has come to light that the U.S. military came close to attacking a Chinese vessel in Middle Eastern waters based on an erroneous intelligence assessment generated by artificial intelligence (AI). Details of what CNN reported on September 18, 2026, are covered in depth in reporting by Engadget’s staff@engadget.com (Ian Carlos Campbell). Had the information not been verified in time, it could have developed into an armed clash, and the case is drawing attention as an example illustrating the dangers of AI adoption at the Department of Defense.

According to Engadget, the U.S. military had advanced to the point of being just about to seize the vessel. As testimony from a source, the following words describing the situation were quoted.

“almost started a war”

The Department of Defense has not offered an official view to Engadget at this time. Engadget has asked the Department of Defense for comment and says it will update its article once a response is received.

Crisis Triggered by an AI-Generated Report

The assessment in question was reportedly prepared by an analyst assigned to the special operations forces. The analyst is said to have asked an AI chatbot about the cargo manifest of the Chinese vessel and compiled an assessment based on its output. According to sources who spoke to CNN, the chatbot combined open-source information with secret signals intelligence and concluded that the vessel was carrying components of a nuclear weapons program.

The judgment as stated in the original was reportedly as follows.

“components of a nuclear weapons program”

“open-source intelligence with secret signals intelligence”

Based on this assessment, the U.S. military moved forward with procedures to seize the vessel. During verification, however, it was discovered that part of the assessment had been generated by AI, and the action was stopped just before it was taken. CNN has not been able to determine what the vessel was actually carrying. It also remains unclear whether the AI platform used by the analyst was a commercial product or a system designed for government use.

The fact that open-source and classified information were mixed without distinction in the process of preparing the assessment aggravated the situation. AI output is often well-written and appears to have sources. As a result, on busy front lines there is a structural tendency to skip checking primary sources. This case showed that this structural weakness directly affected a military decision.

The Pentagon’s AI-First Policy and Its Background

This incident occurred in the midst of the Department of Defense’s push for organization-wide AI adoption. In January 2026, Secretary of Defense Pete Hegseth issued a memorandum directing the transformation of the Department of Defense into an AI-first organization. It reportedly called for actively testing platforms provided by major U.S. AI companies and accelerating their application to missions.

The policy covers a wide range of areas, including intelligence analysis, operational planning, and logistics support. The ability to organize large volumes of documents, signals, and images in a short time is highly attractive to a defense organization facing personnel shortages and growing volumes of information. On the other hand, there is a strong view that verification systems to ensure output accuracy and institutional frameworks concerning responsibility have not kept pace.

The fact that the specific AI platform used has not been identified is also fueling debate. If it was a commercial conversational AI, inaccuracies in training materials and citations would be a problem. Even with a system tailored for government use, the possibility remains that errors could be introduced at the retrieval-augmentation or summarization stage. In either case, the effectiveness of a system in which humans bear final decisions has been called into question.

Dispute Over Military Use of Commercial Models

Relations with private AI companies surrounding the Department of Defense’s AI-first policy also involve tensions. Anthropic has refused to allow its platform to be used for the development of autonomous weapons, bringing its conflict with the government to the surface. It was reportedly at one point subject to a government ban on use. Differences of opinion over safety and the boundaries of military use lie behind the dispute.

Meanwhile, other companies are expanding their cooperation with the Department of Defense. In February 2026, SpaceXAI reportedly signed a contract to make its conversational AI, Grok, available for military use. In May 2026, NVIDIA, Microsoft, and Amazon reportedly formed partnerships to provide AI technology to the Department of Defense. Because government contracts are large and promise long-term revenue, they hold strategic significance for major companies.

Reporting by Engadget’s staff@engadget.com (Ian Carlos Campbell) points out the gap between the incentives and the dangers of such contracts. Government deals are highly profitable, so companies rush to enter. However, malfunctions in the military domain can cause damage incomparable to bugs in ordinary business software. How to balance convenience and safety is also being asked of suppliers.

Dangers of AI Use in the Military Domain

In the field of military intelligence, even a slight error can lead directly to a decision to use force. When the presence of nuclear-related materials is suggested, as in this case, the chain of command is required to respond swiftly. In situations where the time needed for verification conflicts with demands for immediate response, polished AI output risks creating false confidence.

Particularly problematic is the conflation of open-source and classified information. Secret signals intelligence is normally strictly managed together with its collection methods and reliability. If AI combines the two without distinction, the weighting of sources collapses. Users may receive only a plausible-sounding conclusion and lose sight of the hierarchy of evidence.

Verification of AI-generated output is a common challenge on development sites as well. For example, as GitLab’s high-speed Git for AI agents, “Project Switch” shows, the faster automation becomes, the more important mechanisms for tracking change history and verification become. Many observers argue the same applies to military intelligence, and that recording the generation process and re-checking by humans must be built in as institutional requirements.

In this case, fortunately the AI-generated nature was discovered before execution. However, details about how it was discovered and the role of the deployment responsible for verification have not been reported in detail. Concrete measures such as mandatory disclosure of AI use, preservation of citations, and independent double-checks are essential to prevent recurrence. It could be said that operational challenges that cannot be solved by improved technical performance alone have come to light.

Editorial Opinion

On short-term impacts. Over the next three to six months, the Department of Defense is expected to establish interim guidelines on the use of generative AI in intelligence analysis. Operational rules may be clarified, such as requiring disclosure of AI involvement in outputs and prohibiting mixing with classified information. On the supplier side, conditions for use of military-oriented features are also likely to be reviewed, increasing accountability for contract terms.

From a long-term perspective, verification infrastructure for military AI is expected to be developed over the next one to three years. Designs premised on recording generation processes and tracking citations could be incorporated into procurement requirements. Rather than relying solely on improvements to commercial platforms, building independent evaluation systems to detect errors is likely to become a market focus.

Questions from our editors. The question being asked is whether the problem lies in AI performance or in human operations. Can the structure in which verification work is skipped due to busy front lines be resolved by technology alone? How to design systems premised on errors will be central to future debate.

References

• “AI almost led the US military to start a war with China, report says”, by staff@engadget.com (Ian Carlos Campbell) — Engadget, 2026-09-18T22:44:45.000Z (ARR)
• Source URL: https://www.engadget.com/2263043/ai-almost-led-the-us-military-to-attack-china-report-says/