The mechanism half of this review rested on two claims about machine-generated assertion: that models are optimised in a way that rewards confident guessing over admitting uncertainty, and that a model's confidence is the wrong kind of information for deciding whether to act. Both deserve more than a supporting citation each. This entry surveys the trustworthy-machine-learning field those claims live in, and, more importantly, tests the load-bearing one against the field's mainstream view rather than against the single paper that happens to suit it. The test is uncomfortable, and it is the point.
Calibration is real, and largely fixable
Start with confidence. The foundational result here is that modern neural networks are poorly calibrated: their expressed confidence does not track their accuracy, and the effect grew as networks got deeper and wider (Guo et al., 2017). That corroborates, at a more general level, the specific finding about language models the review already cited (Groot and Valdenegro-Toro, 2024). But the same work shows something more useful for this argument: the miscalibration is substantially repairable after the fact, by a single-parameter adjustment known as temperature scaling, which is described as surprisingly effective across most datasets (Guo et al., 2017).
That calibration yields to so simple a fix is not a weakness in the argument. It is the argument. If the problem with model confidence were that it is badly calibrated, then a one-parameter correction would largely solve it, and there would be no deeper issue to discuss. The claim this work makes is different and survives good calibration: confidence, however well tuned, is a statement about the generator's own internal distribution, not evidence that the asserted condition holds in the world. A perfectly calibrated model that reports high confidence has still observed nothing. Calibration improves the odds; it does not supply a basis. Being able to cite that calibration is fixable is what lets the review say this without hand-waving.
The mainstream view disagrees with the claim the argument leaned on
The harder test concerns hallucination. The review's mechanism argument leaned on the position that hallucination is a structural, more or less inevitable consequence of how models are trained and evaluated: under a grading scheme that gives no credit for saying "I do not know," confident guessing dominates as a scoring strategy (Kalai et al., 2025). That is a clean, strong claim, and it does a lot of work.
It is also not the mainstream view. The major surveys of hallucination in language models take a different stance, and they agree with each other while cutting the phenomenon in different ways, which is what makes the agreement worth trusting. One, peer-reviewed, splits hallucination into failures of factuality and failures of faithfulness (Huang et al., 2023); another, independently, into claims that conflict with the user's input, with the preceding context, or with established fact (Zhang et al., 2023). Different taxonomies, same underlying account. It offers a careful taxonomy, splitting hallucination into failures of factuality and failures of faithfulness, and it attributes the phenomenon to specific, addressable causes across the data, training, and inference stages. Its whole orientation is toward mitigation: better data curation, retrieval augmentation, model editing, and refined decoding, all aimed at driving the rate of hallucination down toward acceptable levels. It treats hallucination as a bug to be reduced, not as an intrinsic and permanent feature. Taken at face value, that sits awkwardly against the inevitability claim, and a review that cited only the convenient paper and ignored the survey would deserve the suspicion it invited.
Why the argument survives the disagreement
The tension is real, and it resolves once the two positions are seen to answer different questions. The mitigation literature asks whether the rate of hallucination can be reduced. The answer is yes, and its evidence for that is good. The structural argument asks whether the incentive to assert under uncertainty can be removed. The answer, for as long as evaluation gives abstention no credit, is no. These are compatible. The rate of false assertion can fall while the underlying property persists: a model still emits confident claims whose confidence is uncorrelated with their basis, simply fewer of them.
And for this work, the rate was never the variable that mattered. A thinner stream of unmarked claims is not a safer one, because the receiver still has no way to tell which of the remaining claims is the one without a basis. Suppose mitigation succeeds beyond all current expectation and drives the hallucination rate to a fraction of a percent. The receiver acting on a claim still cannot know whether the claim in hand is in the ordinary ninety-nine percent or the dangerous fraction, because nothing about the claim marks its basis. The problem this work addresses is the absence of that mark, and no amount of rate reduction supplies it. So the mainstream mitigation view is not a refutation to be overcome. It is a position the argument comfortably accommodates, because the argument was never a claim about how often models are wrong. It is a claim about the missing signal that would let a receiver tell proven from asserted, whatever the rate.
Stating this plainly has a cost and a benefit. The cost is giving up the strongest possible phrasing, that hallucination is simply inevitable, which the literature does not support. The benefit is that the argument no longer depends on that contested claim at all. It needs only two things, both uncontested even by the optimists: that basis-free assertion continues to occur, and that nothing in the pipeline marks it. The first is conceded by everyone; the second is the gap itself.
A constructive thread worth following
One strand of this field points forward rather than back. Alongside the work on why models assert too freely, there is a body of research on teaching models to abstain: selective prediction, where a model may decline to answer rather than guess (Geifman and El-Yaniv, 2019), and learning to defer, where it hands the decision to a downstream expert (Mozannar and Sontag, 2020). This is, in effect, the machine-learning field building its own partial answer to the incentive problem, a model that can say "not this one, ask someone else." It is a natural neighbour to the idea at the centre of this work, where a claim's admissibility for a given action is exactly the question of whether it should be acted on or deferred. That connection is noted here and taken up later, where the discipline itself is built.
References
Geifman, Y. and El-Yaniv, R. (2019). SelectiveNet: A deep neural network with an integrated reject option. Proceedings of the 36th International Conference on Machine Learning. arxiv.org/abs/1901.09192
Groot, T. and Valdenegro-Toro, M. (2024). Overconfidence is Key: Verbalized Uncertainty Evaluation in Large Language and Vision-Language Models. arXiv preprint arXiv:2405.02917. arxiv.org/abs/2405.02917
Guo, C., Pleiss, G., Sun, Y. and Weinberger, K. Q. (2017). On calibration of modern neural networks. Proceedings of the 34th International Conference on Machine Learning, pp. 1321-1330. arxiv.org/abs/1706.04599
Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B. and Liu, T. (2023). A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems. arxiv.org/abs/2311.05232
Kalai, A. T., Nachum, O., Vempala, S. S. and Zhang, E. (2025). Why Language Models Hallucinate. arXiv preprint arXiv:2509.04664. arxiv.org/abs/2509.04664
Mozannar, H. and Sontag, D. (2020). Consistent estimators for learning to defer to an expert. Proceedings of the 37th International Conference on Machine Learning. arxiv.org/abs/2006.01862
Zhang, Y., Li, Y., Cui, L., Cai, D., Liu, L., Fu, T., Huang, X., Zhao, E., Zhang, Y., Xu, C., Chen, Y., Wang, L., Luu, A. T., Bi, W., Shi, F. and Shi, S. (2023). Siren's Song in the AI Ocean: A Survey on Hallucination in Large Language Models. arXiv preprint arXiv:2309.01219. arxiv.org/abs/2309.01219