The readings so far established a gap: no existing mechanism lets a receiving owner determine whether the evidence behind a claim is sufficient for the action they intend to take. But a gap only matters if the failure it leaves open is real and persistent. This is the other half of the review, and its job is to show that False Determinism is not a passing quirk of today's systems but a structural failure, manufactured at the source, uncorrectable by confidence, invisible to inspection, and worsened in transit. Four findings, already established in the literature, assemble into that case with no weak link.
Generation produces basis-free assertion by design
The load-bearing point, and the one that makes the failure structural rather than incidental. Recent work on why language models hallucinate establishes that the behaviour is not a defect awaiting a fix but a predictable product of how models are trained and evaluated. Hallucinations, the argument runs, originate simply as errors in binary classification, and models are optimised to be good test-takers, for whom guessing when uncertain improves the score (Kalai et al., 2025). Under any grading regime in which saying nothing earns zero, confident guessing strictly beats admitting ignorance.
The consequence is decisive. A correctly-functioning, well-trained model asserts under uncertainty by design, because the entire training and evaluation ecosystem rewards it for doing so. This lifts the problem from the familiar observation that AI sometimes gets things wrong, which is contingent and might be fixed, to something stronger: the ecosystem systematically produces claims whose confidence is uncorrelated with their basis. Better models do not remove the failure. The failure is what the target rewards.
Fairness to the field requires a caveat here, and it turns out to sharpen the point rather than blunt it. The mainstream view of hallucination, represented by the major peer-reviewed surveys, is not that it is inevitable but that it is mitigable: better training data, retrieval augmentation, and refined decoding can drive the rate of false assertion down toward acceptable levels (Huang et al., 2023). That view is well evidenced, and it appears to sit uneasily beside the claim that the failure is structural. The tension resolves once the two are seen to answer different questions. The mitigation literature asks whether the rate of hallucination can be reduced, and the answer is yes. The structural argument asks whether the incentive to assert under uncertainty can be removed, and the answer, while evaluation gives abstention no credit, is no. These are compatible: the rate can fall while the property persists. And for this work the rate was never the load-bearing variable. A stream of unmarked claims is no safer for being a thinner stream, because the receiver still cannot tell which of the survivors lacks a basis. The argument here needs only that basis-free assertion continues to occur, which even the optimists concede, and that nothing in the pipeline marks it, which is the gap itself.
Confidence is the wrong category of evidence
A model's expressed confidence is poorly calibrated in practice: the evidence shows language and vision-language models to have high calibration error and to be overconfident most of the time (Groot and Valdenegro-Toro, 2024). But this empirical fact is not where the argument rests, and the reason is worth being exact about. If poor calibration were the whole problem, the remedy would be better calibration, and the problem would be temporary.
The claim that survives is stronger and does not depend on the current state of the art. Confidence, however well calibrated, is the wrong kind of information. It is a statement about the generator's own internal distribution, not evidence that the asserted condition holds in the world. A perfectly calibrated model asserting something at high confidence has still observed nothing; it has reported the shape of its own uncertainty with commendable honesty. Calibration improves the odds. It does not supply a basis. And because the deficiency is one of category rather than quality, the remedy cannot belong to the model vendor. It belongs to the system that decides what may be acted upon.
An absence, not a concealment
There is a philosophical account of exactly this indifference. Frankfurt's analysis of assertion made with no regard for the truth has been applied directly to language models (Hicks et al., 2024), and it names, at the level of the utterance, the same property observed here: a generated value bears no relation to its own truth. Not a broken relation, not a sincere one, none at all.
The concern of this work is downstream of that, and more mechanical. The question is not what kind of utterance a generated claim is, but what happens when a value with no relation to its own truth reaches a consumer that can only treat it as settled. The philosophy names the property; this work traces its consequence at an interface. That is also why the failure cannot be met by detection: there is nothing concealed in the claim to be found. The deficiency is not hidden in it. It is absent from it.
The transfer layer strips what little remains
The last link is handover itself. In the most intensively studied setting, clinical handover, the characteristic failure is omission: information is simply not carried across, and absent information presents nothing to correct, while a discrepancy at least offers something to catch (Yeom et al., 2026). Transfer between owners loses information as a matter of course.
One limit must be stated plainly, because overclaiming it would be the very failure this work is about. That literature measures the loss of facts, not the loss of basis, and was not designed to measure the latter; patient harm is the outcome it tracks, and whether a fact was transferred is the variable it can measure. That handover strips the epistemic standing of what survives, how firmly a thing was known, is an argument this work must make, not a finding it can borrow. And yet the concession is itself telling. A mature, safety-critical, heavily-researched discipline converged on standardising what is handed over, through structured forms and checklists, and never once on standardising how firmly it is known. That silence is the strongest available sign that the gap is real, and that it went unnoticed only because, until the originator could be a generator rather than an observer, it did not need to be noticed.
The mechanism, assembled
The four findings compose into a single claim with no weak link. Generation produces basis-free assertion, by the design of its incentives. The confidence it emits is the wrong category of information to repair the deficit, calibrated or not. The deficit is a genuine absence of any relation to truth, not a hidden error to be detected. And the transfer layer that carries the result strips what little basis-signal there was, as handover demonstrably does with information in general.
So the failure is manufactured at the source, uncorrectable by confidence, invisible to inspection, and stripped further in transit. False Determinism is structural at every link. It requires no negligence, no adversary, and no model error; it occurs with every part working as designed. That is exactly why the gap matters: the thing it leaves un-adjudicated is not rare or accidental. It is produced continuously, at scale, by systems all functioning correctly.
References
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
Hicks, M. T., Humphries, J. and Slater, J. (2024). ChatGPT is bullshit. Ethics and Information Technology, 26(2), 38. eprints.gla.ac.uk/327588
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
Yeom, S., Kim, M.-G. and Park, J. H. (2026). Understanding nursing handoff errors in clinical practice: trends and contributing factors based on a systematic review and meta-analysis. BMC Nursing, 25, 359. doi.org/10.1186/s12912-026-04607-x