🇰🇷South Korea·AI News·19 Nov 2026·via Korea It Times

    Jung’s In-Depth Report Proposes Standards for Verifying AI Hallucinations

    Manripungchwi Jung Yeon-tae, President of the Korea Information Service of the Construction Industry (KCINET), published an In-Depth Analysis Report on Generative AI Errors and Hallucination Verification on November 18, 2026. The report proposes new standards for assessing generative AI reliability, advocating for scrutiny of evidence, verification steps, and final output, rather than relying on single accuracy or hallucination scores.

    The Korea Information Service of the Construction Industry (KCINET), led by President Jung Yeon-tae, argues that current methods for evaluating generative artificial intelligence (AI) fall short by relying on single accuracy or hallucination scores. Jung's November 18, 2026, In-Depth Analysis Report on Generative AI Errors and Hallucination Verification highlights that AI systems can invent facts, exaggerate source material, or falsely claim to have performed verification checks. A credible assessment, the report suggests, must therefore examine the evidence supporting an AI's answer, the actual verification steps undertaken, and the final output.

    The report combines a narrative literature review with proposed evaluation and system designs, without presenting new experiments comparing AI models. It reviews existing research on hallucinations, reasoning failures, and verification techniques, focusing on how AI should respond to missing or contradictory evidence and changes in underlying source material. The proposed test questions and the OXMV framework discussed in the report remain proposals that require empirical validation.

    Jung's report clarifies that fabricating a non-existent paper differs from an arithmetic error, and combining such distinct error types into a single score can obscure the nature of the failure. It also notes that even a hallucination rate is highly dependent on the unit of measurement. For instance, five errors across 100 claims could represent a 5% claim-level error rate, but if those errors occur in five different responses, 50% of responses contain an error. The report cites MDH Bench figures, such as 13.30% for Claude Sonnet 4 and 14.45% for GPT 5.2, but cautions that these were under specific testing conditions and should not be treated as universal rates or absolute rankings.

    Furthermore, the report challenges the assumption that longer explanations or extended reasoning traces inherently guarantee accuracy. It references GSM-Symbolic research, which found that changes in mathematical question parameters could affect model performance, and Apple’s 2025 The Illusion of Thinking, which investigated reasoning model limits. Other research noted that models did not always disclose the influence of hints, and self-correction without external feedback could be difficult. These findings support checking calculations, source evidence, and actual verification outcomes rather than relying on the apparent sophistication of an AI's explanation.

    Even advanced techniques like retrieval-augmented generation (RAG) and fact-checking tools have limitations. Missing or unsuitable retrieved material can still lead to errors, and an AI might faithfully reflect an incorrect source or exaggerate an otherwise accurate one. The report examines various approaches like knowledge-graph-based revision (KGR), Self-RAG, corrective retrieval-augmented generation (CRAG), KnowHalu, and Chain-of-Verification (CoVe), which target different stages of evidence assessment and checking. It also reviews semantic-entropy methods for detecting uncertain generations, while observing that consistently repeated wrong answers may evade such detection.

    Share this article

    Original reporting by Korea It TimesWe don't republish, read the full story →

    Related reading

    6 stories