Abstract

Large language model (LLM) hallucination research has produced a rich but fragmented picture: many papers propose a cause, many others propose a detector, and few connect the two systematically. This paper makes two original contributions on top of a narrative review of the field, and strengthens both against external evidence rather than resting on the author's judgment alone. First, it introduces a Cause–Detection Coverage Framework: a reasoned nine-by-six matrix scoring, for each of nine documented hallucination-cause categories, how well each of six detection-technique families can in principle catch it. The framework is then checked retrodictively against four independent, previously published empirical findings; three corroborate its predictions, and one — a 2026 finding that white-box activation probing also collapses to chance on confidently fabricated knowledge-gap content — prompted a documented, evidence-driven revision of one cell rather than being set aside. The surviving pattern is a structural blind spot that individual papers hint at but do not quantify: the cheapest, most widely deployed black-box detectors — consistency sampling and token-probability uncertainty — systematically underperform on exactly the lifecycle-instilled causes (training-data bias, knowledge overshadowing, instruction-tuning error propagation, RLHF-induced sycophancy) that produce confidently and consistently wrong output. Second, the paper presents an original empirical analysis of the growth trajectory of documented AI-hallucination court cases, built from eleven dated, cross-corroborated public snapshots (including one primary-sourced, author-queried count) of a legal-incident tracker spanning December 2023 to September 2026. Beyond a single exponential fit, four competing growth models (exponential, linear, power-law, logistic) are formally compared by AIC and BIC; a logistic (saturating) model is preferred by a wide margin (ΔAIC ≈ 41 over the exponential), and bootstrap confidence intervals are reported for the exponential growth rate and doubling time. Both contributions remain explicitly bounded: the coverage framework is now partially, not fully, validated against external evidence, and the growth analysis measures documentation and enforcement activity, not the underlying LLM hallucination rate. The paper's broader narrative review of causes, benchmarks, and mitigation techniques, and its critical discussion of the field's evidence base, are retained to situate both contributions.

References

• Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., & Liu, T. (2025). A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2), 1–55.
• Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y., Chen, D., Dai, W., Chan, H. S., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1–38.
• Zhao, W. X., Zhou, K., Li, J., et al. (2023). A survey of large language models. arXiv:2303.18223.
• Lin, S., Hilton, J., & Evans, O. (2022). TruthfulQA: Measuring how models mimic human falsehoods. arXiv:2109.07958.
• Li, J., Cheng, X., Zhao, W. X., Nie, J.-Y., & Wen, J.-R. (2023). HaluEval: A large-scale hallucination evaluation benchmark for large language models. arXiv:2305.11747.
• Min, S., Krishna, K., Lyu, X., Lewis, M., Yih, W., Koh, P. W., Iyyer, M., Zettlemoyer, L., & Hajishirzi, H. (2023). FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. arXiv:2305.14251.
• Thorne, J., Vlachos, A., Christodoulopoulos, C., & Mittal, A. (2018). FEVER: A large-scale dataset for fact extraction and verification. Proceedings of NAACL-HLT.
• Ravi, S., et al. (2024). HaluBench: A hallucination evaluation benchmark across real-world domains. arXiv:2407.00499.
• Muhlgay, D., Ram, O., Magar, I., et al. (2024). Generating benchmarks for factuality evaluation of language models (FACTOR). arXiv:2307.06908.
• Manakul, P., Liusie, A., & Gales, M. J. F. (2023). SelfCheckGPT: Zero-resource black-box hallucination detection for generative large language models. arXiv:2303.08896.
• Mündler, N., He, J., Jenko, S., & Vechev, M. (2023). Self-contradictory hallucinations of large language models: Evaluation, detection and mitigation. arXiv:2305.15852.
• Azaria, A., & Mitchell, T. (2023). The internal state of an LLM knows when it's lying. arXiv:2304.13734.
• Ganesh, P., Shokri, R., & Farnadi, G. (2026). Rethinking hallucinations: Correctness, consistency, and prompt multiplicity. arXiv:2602.00723. To appear at EACL 2026.
• Farquhar, S., Kossen, J., Kuhn, L., & Gal, Y. (2024). Estimating the hallucination rate of generative AI. arXiv:2406.07457.
• Gekhman, Z., Yona, G., Aharoni, R., Eyal, M., Feder, A., Reichart, R., & Herzig, J. (2024). Does fine-tuning LLMs on new knowledge encourage hallucinations? arXiv:2405.05904.
• Hu, M., He, B., Wang, Y., Li, L., Ma, C., & King, I. (2024). Mitigating large language model hallucination with faithful finetuning. arXiv:2406.11267.
• Ying, S., et al. (2024). Knowledge overshadowing causes amalgamated hallucination in large language models. arXiv:2407.08039.
• Xie, Z., & Sun, Z. (2026). Mitigating hallucination on hallucination in RAG via ensemble voting (VOTE-RAG). arXiv:2603.27253.
• Zhang, Q., Zhang, H., Pang, L., Zheng, H., & Zheng, Z. (2026). Stable-RAG: Mitigating retrieval-permutation-induced hallucinations in retrieval-augmented generation. arXiv:2601.02993.
• Do, N., et al. (2025). Finetune-RAG: Fine-tuning language models to resist hallucination in retrieval-augmented generation. arXiv:2505.10792.
• Banerjee, S., Agarwal, A., & Singla, S. (2024). LLMs will always hallucinate, and we need to live with this. arXiv:2409.05746.
• Kim, Y., Jeong, H., Chen, S., Li, S. S., Lu, M., Alhamoud, K., Mun, J., et al. (2025). Medical hallucination in foundation models and their impact on healthcare. arXiv:2503.05777.
• Asgari, E., Montaña-Brown, N., Dubois, M., Khalil, S., Balloch, J., Au Yeung, J., & Pimenta, D. (2025). A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation. npj Digital Medicine, 8(1), 274.
• Wille, J. (2025, May 29). Lawyer sanctioned $6,000 for AI-generated fake legal citations. Bloomberg Law.
• The Register. (2023, June 22). Lawyers who cited fake cases hallucinated by ChatGPT must pay. Reporting on Mata v. Avianca Airlines, S.D.N.Y.
• Yahoo News / AP. (2025). AI hallucinations in court documents are growing. Reporting on the AI Hallucination Cases Database case counts by year (2023–2025).
• Charlotin, D. (2023–2026). AI Hallucination Cases Database [dataset]. damiencharlotin.com/hallucinations.
• LawSites / The Law GPT. (2026). AI hallucination sanctions: A lawyer's guide for 2026.
• Wan, R., et al. (2026). “My AI is lying to me”: User-reported LLM hallucinations in AI mobile app reviews.
• Saxena, R., & Bhattacharyya, P. (2024). A survey on hallucination detection methods: Intrinsic and extrinsic taxonomies. As cited in Huang et al., 2025.
• Platinum IDS Blog. (2026, April). 1,227 fabricated citations and counting: Inside the AI hallucination crisis hitting courts worldwide. [Charlotin database snapshot, early April 2026: 1,227 cases.]
• HAQQ.ai Blog. (2026, June 9). AI hallucination cases: The 1,598-case sanctions tracker. [Charlotin database snapshot, 9 June 2026: 1,598 cases.]
• Legal AI Space. (2026, July 2). 1,600+ AI hallucination cases: What every law firm should learn. [Charlotin database snapshot, 2 July 2026: 1,668 cases.]
• GC.ai Blog. (2026). AI hallucination legal cases: A sanctions tracker. [Charlotin database snapshot, 17 July 2026: 1,769 cases.]
• Suprmind.ai. (2026, August 27). AI hallucination rates, statistics & benchmarks in 2026. [Charlotin database snapshot, 27 August 2026: 1,963 cases; also reports the 719-case January 2026 and 1,227-case early-April 2026 figures.]
• The Law GPT. (2026, July). AI hallucination sanctions: A lawyer's guide for 2026. [Corroborating snapshot, ≈1,490 cases, May 2026.]
• Volokh, E. (2025, May 18). “AI Hallucination Cases,” from courts all over the world. Reason (The Volokh Conspiracy). [Independent legal-commentary snapshot: 87 cases.]
• Willison, S. (2025, May 25). AI Hallucination Cases. simonwillison.net. [Independent technology-commentary snapshot citing the Charlotin tracker directly: 116 cases, 12 countries.]
• Manach, J.-M. (2025, July 1). Il génère des plaidoiries par IA, et en recense 160 ayant « halluciné » depuis 2023. Next (next.ink). [Independent technology-press snapshot: ≈160 cases.]
• Charlotin, D. (2026, September 5). AI Hallucination Cases Database [dataset, live query]. damiencharlotin.com/hallucinations. [Primary-source snapshot: 2,021 cases.]
• Zhang, J., Li, X., Yu, K., Peng, B., & Chen, Y. (2023). SAC3: Reliable hallucination detection in black-box language models via semantic-aware cross-check consistency. arXiv:2311.01740.
• Bhatt, M. (2026). Case study: Predictive coding and information bottleneck for hallucination detection in large language models. arXiv:2601.15652. [Single-author preprint, no confirmed peer-reviewed venue at time of citation; the sycophancy–consistency-check claim attributed within it to “Fadeeva et al.” was not independently verified by the present author and is treated as weak corroboration accordingly, per Section 6.1.]
• Fadeeva, E., Rubashevskii, A., Shelmanov, A., Petrakov, S., Li, H., Mubarak, H., Tsymbalov, E., Kuzmin, G., Panchenko, A., Baldwin, T., Nakov, P., & Panov, M. (2024). Fact-checking the output of large language models via token-level uncertainty quantification. In Findings of the Association for Computational Linguistics: ACL 2024 (pp. 9367–9385). [Cited in Section 5–6 for token-level uncertainty quantification generally; not the source of the sycophancy-specific claim in
• , which traces to a different, unverified citation.]
• Chauhan, M. S. (2026). DECK: A consistency × confidence taxonomy of LLM hallucinations. arXiv:2606.02289.