Definition

AI hallucination occurs when a generative model confidently outputs information not grounded in its inputs or training data — inventing facts, misreading structured data, or fabricating citations. In high-stakes domains the failure mode is not random noise but plausible, authoritative-sounding errors.

Key Points

  • 2026-09-18: Documented near-miss — US Special Operations Command analyst’s chatbot hallucinated nuclear components on a Chinese vessel manifest; aircraft airborne before abort (2026-09-18-ai-hallucination-us-military-operation, 2026-09-19-ai-hallucination-us-military-cnn)
  • Analyst compounded error by using AI again to format hallucinated output into standard intelligence report format
  • Source described report as “entirely false” and episode as almost starting a war
  • Chatbot vendor/model identity not publicly disclosed as of reporting date

Near-miss framing only — operation aborted before kinetic action. Do not claim a strike was ordered or executed.

Sources