Hallucination Detection — References

The taxonomy of the defect, the reference-free detection methods, and the evidence about whether a model's own confidence is worth anything.

The defect

• Survey of Hallucination in Natural Language Generation — Ziwei Ji, Nayeon Lee, Rita Frieske and colleagues, ACM Computing Surveys, 2023. Cited here for the intrinsic and extrinsic split: output contradicting its input and output asserting something no input mentioned are different defects, and only the second is what this capability faces when there is no input to contradict. Cited by name.
• TruthfulQA: Measuring How Models Mimic Human Falsehoods — Stephanie Lin, Jacob Hilton and Owain Evans, Annual Meeting of the Association for Computational Linguistics, 2022. The benchmark built around questions where the plausible answer is a common misconception, and the finding that larger models were in some cases less truthful on them. The cleanest demonstration that fluency and accuracy move independently. Cited by name.
• Why Language Models Hallucinate (2025) — the argument that evaluation regimes rewarding a guess over an abstention produce models that guess, which locates part of the cause in how systems are scored rather than in how they are trained. Cited by name.

Detecting without a reference

• SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models — Potsawee Manakul, Adian Liusie and Mark J. F. Gales, Conference on Empirical Methods in Natural Language Processing, 2023. Sampling several responses and measuring their mutual consistency, with no access to model internals and no external corpus. The reference method for this family and the clearest statement of its assumption, which is that fabricated content varies between samples. Cited by name.

Whether the model knows

• Language Models (Mostly) Know What They Know — Saurav Kadavath and colleagues at Anthropic,
  2022. Evidence that models can be trained to predict whether their own answers are correct, along with the limits of that self-assessment as tasks move away from the calibration distribution. Cited by name.
• Teaching Models to Express Their Uncertainty in Words — Stephanie Lin, Jacob Hilton and Owain Evans, Transactions on Machine Learning Research, 2022. Verbalised confidence as a trainable behaviour, and the comparison with log-probability-derived confidence that is the basis for choosing between them. Cited by name.

No verified reference is recorded for referent resolution against external authorities as a detection method, for consequence-weighted verification routing, or for the effect of visible uncertainty markers on user trust over time. The published work detects hallucination; almost none of it measures what a product should then do.