Pith. sign in

REVIEW 3 major objections 3 minor 40 references

Machine Mirages: Defining the Undefined

T0 review · 3 major / 3 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that machine failure modes like hallucination and confabulation can be formally defined as risk quantities, and that a model keeping all 34 such risks low is guaranteed to generalize to aligned downstream tasks.

desk verdict A useful taxonomy buried under a broken risk guarantee: Theorem 6 does not follow, so the paper's central quantitative claim should not be trusted. read the letter →

arxiv 2506.13990 v1 pith:Y7HTEQDR submitted 2025-06-03 cs.AI

Hamidou Tembine This is my paper
classification cs.AI
keywords machinemiragescomputationalcognitivepathologieshallucinationformalizationexpectilevalue-at-riskmulti-objectiveriskgenerativepathologydiscriminativemean-field-typegames
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper's central claim is that the loosely used vocabulary of AI failure — hallucination, confabulation, delusion, semantic drift, bias amplification, calibration failure, and 28 other named failures — can be turned into precisely defined, mathematically grounded pathologies: each one is given an operative formula, and each pathology's risk is quantified by an expectile value-at-risk. If those definitions are accepted, a model that keeps all 34 pathology risks below a small tolerance provably has bounded expected loss on any downstream task whose failure mode is expressible as one of those pathology indicators. This would matter because it converts vague regulatory talk about AI safety into quantitative, checkable conditions: low pathology risk would certify generalization to unforeseen tasks that are structurally aligned with the defined failure modes. The paper also claims that no single pre-trained transformer can be optimal across all 34 risks simultaneously, so risk control is inherently a Pareto-frontier, multi-objective problem managed by a human-machine mean-field-type game.

What carries the argument

The load-bearing machinery is the pairing of 34 pathology indicators (concrete formulas in Tables I and III) with expectile value-at-risk, $R(\text{pathology}_i) = \arg\min_{r} \mathbb{E}[ w_\tau(L_i - r)^2 ]$, where $w_\tau(z) = \tau \mathbf{1}\{z>0\} + (1-\tau)\mathbf{1}\{z\le 0\}$. The indicator turns each qualitative failure label into a binary stochastic event, and the expectile turns that event into an asymmetric, coherent risk number that weights rare catastrophic failures and frequent low-grade degradations separately. Theorem 6's proof then exploits the fact that for a binary indicator the expectile collapses to a monotone transform of the failure probability, giving a linear bound $C \cdot R_i$. Around this core sits the Holonorm transformer architecture — each residual block has the form $(\mathrm{id} + f\circ\mathrm{hn}) \circ (\mathrm{id} + \mathrm{MHA}\circ\mathrm{hn})$ with $\mathrm{hn}(x)=x/(1+\|x\|)$ — which Theorem 4 claims is expressive enough, by universal approximation, to be pre-trained to minimize each risk.

What would settle it

Pick a currently deployed multimodal model, fix one pathology with its Table I or III indicator, collect a corpus where human judges label the indicator, and estimate both the expectile risk and the binary-event rate; if on any pathology the observed ratio of the expectile of the indicator to the pathology risk exceeds the constant asserted in Theorem 6's proof, the claimed bound fails. Alternatively, construct a downstream task whose loss is emotional-persuasion effectiveness regardless of truth, keep all 34 indicator risks low on a test suite, and measure whether the task loss is actually bounded by epsilon; Theorem 7 predicts it will not be.

Watch

Extended reading notes

Core claim

The discovery the paper argues for is that computational cognitive pathologies — machine mirages — are not merely metaphors but are definable stochastic events. For each of 34 named failure modes the paper gives an operational condition: for example, confabulation is 'output that maximizes model probability while scoring below a coherence threshold against a knowledge base,' and referential hallucination is 'output that contains an entity outside both training data and verifiable external reality.' Each is then assigned an expectile value-at-risk, an asymmetric coherent risk measure that weights rare catastrophic failures and frequent low-grade degradations differently. The paper's central technical result, Theorem 6, asserts that if a Holonorm Transformer (a residual attention network whose nonlinearities are the norm-bounded map $\mathrm{hn}(x)=x/(1+\|x\|)$) is pre-trained so that the maximum of the 34 pathology risks is below a tolerance, then for any downstream task whose loss can be written as $\operatorname{ExpVaR}_{\tau}$ of the binary indicator of one of the 34 pathologies, the expected task loss is bounded by a prescribed $\varepsilon$. Corollary 1 makes explicit the flip side: tasks not expressible through these indicators fall outside the guarantee, and Theorem 7 states such structurally unaligned tasks can always be exploited by prompting or chaining.

Load-bearing premise

The generalization guarantee assumes every relevant downstream failure can be written exactly as an expectile risk over one of the 34 predefined pathology indicators; a task whose failure mode is not representable by one of those indicators is outside the bound, and Theorem 7 says such tasks can always be exploited.

Editorial extensions

If this is right

  • If a deployed model keeps all 34 expectile pathology risks beneath a stated tolerance, Theorem 6 certifies bounded expected loss on every downstream task whose failure can be written as expectile risk of one of the 34 indicators, including tasks not seen during training.
  • Because Theorem 1 excludes any transformer that beats all others on every risk, the framework implies any safe-deployment regime must manage a Pareto tradeoff among the 34 risks rather than optimize a single headline metric.
  • Theorem 7 shows post-training governance cannot rely on the model alone: adversarial prompting, chaining, or fine-tuning can always exploit tasks whose objectives lie outside the 34 pathologies, so human oversight and mechanism design remain necessary.
  • The mean-field-type game formulation lets a human prescribe risk thresholds and delegates each machine-intelligence agent to minimize a specific risk under shared compute and data-quality constraints, making the framework directly operational for risk-aware model co-development.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the practical force of Theorem 6 depends on whether real deployment failures decompose into the 34 predefined indicators; a testable extension is to build detectors for each Table I and III formula and empirically estimate the 34 expectile risks on current models, checking whether the claimed bound $C \cdot R_i$ is tight or loose.
  • Editorial inference: the definitions in Tables I and III leave thresholds like $\mathrm{sim} \approx 1$, $\alpha \gg 1$, and $\mathrm{discomfort}(\cdot)$ unspecified; operationalizing those thresholds per application would turn the formal definitions into measurement protocols, and the choice of thresholds may itself change which events count as pathological.
  • Editorial inference: Theorem 4 establishes expressive feasibility, not trainability; without a guarantee that gradient-based pretraining converges to a distribution that simultaneously keeps all 34 risks low, the generalization bound remains conditional on an optimization success that still needs empirical demonstration.
  • Editorial inference: expectile value-at-risk is one of many coherent risk measures; replacing it with CVaR or standard quantile VaR would change the Pareto frontier and the regulatory thresholds, so the choice of asymmetric weighting deserves direct comparison on the same 34 pathology events.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 3 minor

Summary. The paper introduces a taxonomy of 34 "computational cognitive pathologies" (machine mirages), proposes table-based formalizations for generative and discriminative failure modes, and defines an expectile-based risk measure R_i for each pathology. Its central quantitative claim is Theorem 6: if a model keeps all 34 expectile risks below a tolerance, then its expected loss on any downstream task expressible through those pathologies is bounded. The paper also claims the Holonorm Transformer is expressive enough to be pretrained to minimize these risks (Theorem 4), derives a density transformation result for the Holonorm map (Theorem 5), and sketches a mean-field-type game for multi-agent risk control. Corollary 1 limits the generalization guarantee to tasks whose loss is exactly ExpVaR_tau of one of the 34 pathology indicators.

Significance. If Theorem 6 were valid, the paper would offer a quantitative, definition-based bridge from pathology definitions to post-training governance and regulation, which would be a substantial contribution. The paper has genuine strengths: it catalogs a broad range of failure modes, grounds them in concrete African-language audio examples, and explicitly acknowledges in Corollary 1 that its guarantee only applies to structurally aligned tasks. The proof of Theorem 5 (the density of a Holonorm-transformed random vector) is a correct and easily verified application of the change-of-variables formula and the matrix determinant lemma. However, the central theorem is not sound, and the claimed formalizations in Tables I and III are not sufficiently operational to define the quantities on which the risk framework depends.

major comments (3)
  1. [Section III, Theorem 6] The load-bearing step of the proof is false. In Section III, R_i is defined as the expectile of a severity random variable L_i, not of a binary indicator. Theorem 6's proof replaces L_i with I_Path_i in {0,1} and asserts that the expectile of the indicator is bounded above by R_i(h). This assertion fails in general. For example, let L_i = s with probability q and L_i = 0 otherwise, with tau = 0.1. The expectile of L_i is r = tau q s / [(1-tau)(1-q) + tau q]. For any epsilon > 0 and any proposed tolerance delta, set q = 1.5 epsilon and choose s so that r = delta, e.g., s = delta [(1-tau)(1-q) + tau q] / (tau q). Then R_i(h) = delta but the expected binary downstream loss is q = 1.5 epsilon > epsilon, contradicting the claimed implication. The Lipschitz-continuity remark in the proof does not repair the gap, because the map from the severity expectile to the occurrence probability has no uniform Lipschitz constant as s varies. Since Theorem 6 is the paper's central generalization guarantee, this is a fatal internal inconsistency, not merely a missing hypothesis.
  2. [Section II, Tables I and III] The paper repeatedly states that these tables give precise mathematical formalizations, but most entries are informal inequalities or qualitative criteria involving undefined primitives: S_real, E_real, discomfort(.), G, iota(x), S_expert, and unspecified thresholds alpha, gamma, epsilon, delta. As a result, the severity random variable L_i is not actually well defined, and the indicator I_Path_i required by Theorem 6 cannot be derived from the table. This is load-bearing because the expectile risk R_i and the task loss L_T are both defined over these pathology events; without a measurable definition of the event, the theorem's hypothesis is not operational.
  3. [Section III, Theorem 4] The proof of Theorem 4 asserts that "under universal approximation conditions on the attention and linear projections" the Holonorm Transformer can approximate any continuous function. This is not a proof: the required conditions are never stated or verified, and universal approximation of individual functions does not imply simultaneous approximate minimization of 34 different expectile risks, especially when the paper's own Theorem 1 claims these risks are non-aligned and lie on a Pareto frontier. Thus the premise that a model can be pretrained to keep all 34 risks small is unsupported.
minor comments (3)
  1. [Section III, Corollary 1] The claim that the bound holds "if and only if" L_T has the stated expectile form is not proved; Theorem 6 only establishes a sufficient condition under the expressibility assumption. As written, Corollary 1 defines the scope of the guarantee rather than providing an external characterization.
  2. [Throughout] There are several typographical errors, including "c34 omputational" in Theorem 4 and "addd" in the introduction; the manuscript would benefit from a careful proofreading pass.
  3. [Section III, Theorems 1-3] The statements of Theorems 1-3 use different notions of optimality ("outperforms in all risk metrics" versus "minimizes a fixed risk") without defining a common formal sense of optimality, which makes the precise claims difficult to evaluate.

Circularity Check

1 steps flagged · score 8.0 of 10

Theorem 6's 'generalization guarantee' is a definitional tautology: the downstream tasks it covers are defined as expectile losses over the same 34 pathology indicators whose risks are minimized.

  1. self definitional [Section III, Theorem 6 and Corollary 1]
    "Define each downstream taskT by its loss LT (h(x)) := ExpVaRτ(IPathT (h(x))), where IPathT (h(x)) = 1 if the output h(x) exhibits the pathology degrading task T ... The generalization bound ... holds if and only if the loss function LT can be expressed as LT (h(x)) := ExpVaRτ(IPathT (h(x))), where IPathT (h(x)) indicates the occurrence of a pathology from the predefined set of 34 computational pathologies."

    Theorem 6 first defines the target task loss as an expectile over a binary pathology indicator, and Corollary 1 states that the bound holds if and only if LT is expressible in exactly that form over the predefined 34 pathologies. The claimed 'epistemic generalization to all tasks expressible through these pathologies' is therefore a restatement of the definition of the allowed task class: the only tasks covered are those whose loss is a function of the same pathology indicators whose risks are being minimized. The proof's one non-definitional bridge, that the binary-indicator expectile is bounded by the severity-based risk Ri, is asserted without derivation from Section III's definition of Ri over severity Li.

full rationale

The paper's central quantitative claim is Theorem 6, which asserts that bounding the 34 pathology risks bounds the expected loss on any downstream task 'expressible through these pathologies.' Corollary 1 makes the expressibility condition explicit: the bound holds if and only if the task loss is ExpVaRτ(IPathT) for one of the predefined 34 indicators. Consequently, the set of covered tasks is exactly the set of losses constructed from the pathology indicators, so the guarantee restates the task definition rather than providing an independent predictive or empirical result. The proof's would-be bridge from the binary indicator expectile to the severity-based expectile Ri is asserted rather than derived from Section III's definition of Ri over Li; under the binary reading the bound is true by construction, while under the severity reading it is false, since a rare event with tiny severity can make Ri arbitrarily small while the probability of the binary pathology event is large. That mathematical defect is a correctness risk separate from the circularity finding. The mean-field-type game references [36]-[40] include the author's own prior work, but they are used for framing only and are not load-bearing for Theorem 6. No other claimed prediction in the paper reduces to its inputs by construction.

Assumptions & free parameters 8 free parameters · 6 assumptions · 2 invented entities

The central framework assumes the 34 pathologies are well-defined stochastic events and that deployment tasks can be expressed through them; this is what makes Theorem 6's guarantee partly definitional. Definitions rely on hand-chosen thresholds and external knowledge bases, and the Holonorm architecture is proposed without experiments. No data or code is provided, so the ledger is largely conceptual.

free parameters (8)
  • expectile level tau = tau in (0,1), no value given
    The risk measure R(pathology_i) is defined as an expectile value-at-risk at level tau; all risk comparisons depend on this uncalibrated choice.
  • coherence threshold tau_C = unspecified
    Defines confabulation in Table I; different thresholds classify the same output differently.
  • high-probability threshold delta = unspecified
    Used in the Abductive Leap definition and in Theorem 6's task-loss framing; never estimated.
  • exaggeration factor alpha = alpha >> 1
    Defines exaggeration in Table I as a norm inequality relative to y_true; the value of alpha is arbitrary.
  • contextual drift threshold epsilon = unspecified
    Used in the Contextual Drift definition D(c_t, c_{t-k}) >> epsilon; no calibration procedure is given.
  • hypersignification threshold gamma = unspecified
    Defines hypersignification in Table I; no method for choosing gamma is given.
  • trade-off parameter lambda = > 0, unspecified
    In the mean-field-type game cost J_i; balances risk R_i and resource cost, but no value or estimation method is given.
  • number of pathologies = 34
    The taxonomy size is asserted; the game and the theorems count on exactly 34 pathologies without a completeness argument.
assumptions (6)
  • standard math No Free Lunch theorem is applicable to the 34 pathology risk functions and transformer parameter space
    Invoked in Theorem 1; the paper does not verify the task-distribution alignment conditions under which NFL holds.
  • standard math Universal approximation theorem applies to the Holonorm Transformer over compactly supported distributions
    Used in Theorem 4; assumes sufficient width and depth and a smooth bounded activation, but approximation of functions does not imply minimization of expectile risks.
  • domain assumption Each of the 34 pathologies is a well-defined stochastic event with a distribution conditioned on context, user, and task
    Section III states this; without operational definitions of L_i, the expectile risk is not computable.
  • ad hoc to paper Downstream task losses can be expressed as ExpVaR_tau(IPathT(h(x))) for one of the 34 pathologies
    Theorem 6 and Corollary 1 define task loss this way, which makes the generalization guarantee conditional and partly definitional.
  • ad hoc to paper ExpVaR_tau is Lipschitz and bounded by R_i, with L_T <= C R_i
    Used in the proof of Theorem 6; stated without proof and not generally established for expectiles over binary indicators versus severity losses.
  • domain assumption A blockchained, trusted knowledge base E_real exists and is accessible for referential grounding
    Referential hallucination and coherence definitions rely on an external verified fact repository; no implementation or access model is given.
invented entities (2)
  • Holonorm map and Holonorm Transformer block
    purpose: A new bounded normalization and nonlinearity hn(x)=x/(1+||x||) with residual attention and feedforward blocks, claimed in Theorem 4 to be expressive enough to minimize any of the 34 pathology risks.
    No experiments, datasets, or trained models are provided; the universal-approximation argument does not show that risk minimization is achievable with finite data.
  • Computational cognitive pathology (machine mirage) taxonomy of 34 named failures
    purpose: Unified conceptual category for generative and discriminative failures, intended to replace metaphorical terms with formal definitions.
    It is a naming and classification scheme with no falsifiable predictions or external validation; its coverage is not tested against real model outputs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Machine Mirages: Defining the Undefined." pith.science (2026). https://pith.science/paper/Y7HTEQDR

@misc{pith2026250613990,
  author       = {Pith},
  title        = {Pith review of: Machine Mirages: Defining the Undefined},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y7HTEQDR}},
  note         = {Machine review of arXiv:2506.13990}
}
read the original abstract

As multimodal machine intelligence systems started achieving average animal-level and average human-level fluency in many measurable tasks in processing images, language, and sound, they began to exhibit a new class of cognitive aberrations: machine mirages. These include delusion, illusion, confabulation, hallucination, misattribution error, semantic drift, semantic compression, exaggeration, causal inference failure, uncanny valley of perception, bluffing-patter-bullshitting, cognitive stereotypy, pragmatic misunderstanding, hypersignification, semantic reheating-warming, simulated authority effect, fallacious abductive leap, contextual drift, referential hallucination, semiotic Frankenstein effect, calibration failure, spurious correlation, bias amplification, concept drift sensitivity, misclassification under uncertainty, adversarial vulnerability, overfitting, prosodic misclassification, accent bias, turn boundary failure, semantic boundary confusion, noise overfitting, latency-induced decision drift, ambiguity collapse and other forms of error that mimic but do not replicate human or animal fallibility. This article presents some of the errors and argues that these failures must be explicitly defined and systematically assessed. Understanding machine mirages is essential not only for improving machine intelligence reliability but also for constructing a multiscale ethical, co-evolving intelligence ecosystem that respects the diverse forms of life, cognition, and expression it will inevitably touch.

Figures

Figures reproduced from arXiv: 2506.13990 by the authors.

Figure 1
Figure 1. A conceptual diagram illustrating some computational cognitive pathologies in machine co-intelligence. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Some generative pathologies in machine co-intelligence. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 36 canonical work pages

  1. [1]

    Bias in artificial intelligence for medical imaging: fundamentals, detection, avoidance, mitigation, challenges, ethics, and prospects,

    B. Koc ¸ak, A. Ponsiglione, A. Stanzione, C. Bluethgen, J. Santinha, L. Ugga, M. Huisman, M. E. Klontzas, R. Cannella, and R. Cuocolo, “Bias in artificial intelligence for medical imaging: fundamentals, detection, avoidance, mitigation, challenges, ethics, and prospects,” Diagnostic and interventional radiology, vol. 31, no. 2, p. 75, 2025

    work page 2025

  2. [2]

    Addressing bias in artificial intelligence for public health surveillance,

    L. Flores, S. Kim, and S. D. Young, “Addressing bias in artificial intelligence for public health surveillance,” Journal of Medical Ethics , vol. 50, no. 3, pp. 190–194, 2024

    work page 2024

  3. [3]

    Ai racial bias: How text-to-image artificial intelligence generators construct prestigious professions,

    A. M. G ´orska and D. Jemielniak, “Ai racial bias: How text-to-image artificial intelligence generators construct prestigious professions,” in Algorithms, Artificial Intelligence and Beyond . Routledge, 2025, pp. 211–226

    work page 2025

  4. [4]

    Gender bias in natural language processing and computer vision: A comparative survey,

    M. Bartl, A. Mandal, S. Leavy, and S. Little, “Gender bias in natural language processing and computer vision: A comparative survey,” ACM Computing Surveys, vol. 57, no. 6, pp. 1–36, 2025

    work page 2025

  5. [5]

    Hallucination or confabulation? neuroanatomy as metaphor in large language models,

    A. L. Smith, F. Greaves, and T. Panch, “Hallucination or confabulation? neuroanatomy as metaphor in large language models,” PLOS Digital Health , vol. 2, no. 11, p. e0000388, 2023

    work page 2023

  6. [6]

    Memory Helps, but Confabulation Misleads: Understanding Streaming Events in Videos with MLLMs

    G. Zhang, M. Ding, T. Liu, Y . Zhang, and V . Tresp, “Memory helps, but confabulation misleads: Understanding streaming events in videos with mllms,” arXiv preprint arXiv:2502.15457 , 2025

    work page Pith review arXiv 2025

  7. [7]

    Confabulation of different iot approaches with and without data compression,

    A. Sisodia and A. K. Yadav, “Confabulation of different iot approaches with and without data compression,” Comput. Integr. Manuf. Syst. , vol. 28, no. 11, 2022

    work page 2022

  8. [8]

    The cathedral: A jungian framework for artificial general intelligence

    M. Bugay, “The cathedral: A jungian framework for artificial general intelligence.”

    work page

Show all 40 references
  1. [9]

    A review of methods for reducing hallucinations in generative artificial intelligence to enhance knowledge economy,

    Z. Roozbahani, “A review of methods for reducing hallucinations in generative artificial intelligence to enhance knowledge economy,” Knowledge Economy Studies, 2025

    2025

  2. [10]

    Halogen: Fantastic llm hallucinations and where to find them,

    A. Ravichander, S. Ghela, D. Wadden, and Y . Choi, “Halogen: Fantastic llm hallucinations and where to find them,” 2025. [Online]. Available: https://halogen-hallucinations.github.io/

    2025

  3. [11]

    Hallulens: Llm hallucination benchmark,

    Y . Bang, Z. Ji, A. Schelten, A. Hartshorn, T. Fowler, C. Zhang, N. Cancedda, and P. Fung, “Hallulens: Llm hallucination benchmark,” 2025. [Online]. Available: https://arxiv.org/abs/2504.17550

    2025 arXiv

  4. [12]

    Faithbench: A diverse hallucination benchmark for summarization by modern llms,

    F. S. Bao, M. Li, R. Qu, G. Luo, E. Wan, Y . Tang, W. Fan, M. S. Tamber, S. Kazi, V . Sourabh, M. Qi, R. Tu, C. Xu, M. Gonzales, O. Mendelevitch, and A. Ahmad, “Faithbench: A diverse hallucination benchmark for summarization by modern llms,” in Proceedings of the 2025 Conferen...

    2025

  5. [13]

    Longhalqa: Long-context hallucination evaluation for multimodal large language models,

    H. Qiu, J. Huang, P. Gao, Q. Qi, X. Zhang, L. Shao, and S. Lu, “Longhalqa: Long-context hallucination evaluation for multimodal large language models,” 2024. [Online]. Available: https://arxiv.org/abs/2410.09962

    2024 arXiv

  6. [14]

    Benchmarking hallucination in large language models based on unanswerable math word problem,

    Y . Sun, Z. Yin, Q. Guo, J. Wu, X. Qiu, and H. Zhao, “Benchmarking hallucination in large language models based on unanswerable math word problem,” in Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-C...

    2024

  7. [15]

    Ai hallucinations: a misnomer worth clarifying,

    N. Maleki, B. Padmanabhan, and K. Dutta, “Ai hallucinations: a misnomer worth clarifying,” in 2024 IEEE conference on artificial intelligence (CAI) . IEEE, 2024, pp. 133–138

    2024

  8. [16]

    Chatgpt: these are not hallucinations–they’re fabrications and falsifications,

    R. Emsley, “Chatgpt: these are not hallucinations–they’re fabrications and falsifications,” Schizophrenia, vol. 9, no. 1, p. 52, 2023

    2023

  9. [17]

    Hallucination-free? assessing the reliability of leading ai legal research tools,

    V . Magesh, F. Surani, M. Dahl, M. Suzgun, C. D. Manning, and D. E. Ho, “Hallucination-free? assessing the reliability of leading ai legal research tools,” Journal of Empirical Legal Studies , 2024. 14

    2024

  10. [18]

    Ai hallucinations? what about human hallucination?!: Addressing human imperfection is needed for an ethical ai,

    A. Tlili and D. Burgos, “Ai hallucinations? what about human hallucination?!: Addressing human imperfection is needed for an ethical ai,” IJIMAI, vol. 9, no. 2, pp. 68–71, 2025

    2025

  11. [19]

    Estimating the hallucination rate of generative ai,

    A. Jesson, N. Beltran Velez, Q. Chu, S. Karlekar, J. Kossen, Y . Gal, J. P. Cunningham, and D. Blei, “Estimating the hallucination rate of generative ai,” Advances in Neural Information Processing Systems , vol. 37, pp. 31 154–31 201, 2024

    2024

  12. [20]

    The problem of ai hallucination and how to solve it,

    A. Jan ´eafik and O. Dusek, “The problem of ai hallucination and how to solve it,” in European Conference on e-Learning . Academic Conferences International Limited, 2024, pp. 122–128

    2024

  13. [21]

    False responses from artificial intelligence models are not hallucinations,

    S. D. Østergaard and K. L. Nielbo, “False responses from artificial intelligence models are not hallucinations,” Schizophrenia bulletin, vol. 49, no. 5, pp. 1105–1107, 2023

    2023

  14. [22]

    Another reason to call bullshit on ai “hallucinations

    J. Slater and J. Humphries, “Another reason to call bullshit on ai “hallucinations”,” AI & SOCIETY , pp. 1–2, 2025

    2025

  15. [23]

    Siren’s song in the ai ocean: a survey on hallucination in large language models,

    Y . Zhang, Y . Li, L. Cui, D. Cai, L. Liu, T. Fu, X. Huang, E. Zhao, Y . Zhang, Y . Chen et al., “Siren’s song in the ai ocean: a survey on hallucination in large language models,” arXiv preprint arXiv:2309.01219 , 2023

    2023 arXiv

  16. [24]

    A survey of hallucination in large foundation models,

    V . Rawte, A. Sheth, and A. Das, “A survey of hallucination in large foundation models,” arXiv preprint arXiv:2309.05922 , 2023

    2023 arXiv

  17. [25]

    The troubling emergence of hallucination in large language models-an extensive definition, quantification, and prescriptive remediations

    V . Rawte, S. Chakraborty, A. Pathak, A. Sarkar, S. I. Tonmoy, A. Chadha, A. Sheth, and A. Das, “The troubling emergence of hallucination in large language models-an extensive definition, quantification, and prescriptive remediations.” Association for Computational Linguistics, 2023

    2023

  18. [26]

    Legal hallucinations and the adoption of artificial intelligence in the judiciary,

    B. A. Herrera-Tapias and D. H. Guzm ´an, “Legal hallucinations and the adoption of artificial intelligence in the judiciary,” Procedia Computer Science , vol. 257, pp. 1184–1189, 2025

    2025

  19. [27]

    Chatbot confabulations are not hallucinations—reply,

    T. D. Brender, “Chatbot confabulations are not hallucinations—reply,” JAMA Internal Medicine , vol. 183, no. 10, pp. 1177–1178, 2023

    2023

  20. [28]

    Chatbot confabulations are not hallucinations,

    R. Hatem, B. Simmons, and J. E. Thornton, “Chatbot confabulations are not hallucinations,” JAMA Internal Medicine, vol. 183, no. 10, pp. 1177–1177, 2023

    2023

  21. [29]

    Beware of artificial intelligence hallucinations or should we call confabulation?

    H. Berk, “Beware of artificial intelligence hallucinations or should we call confabulation?” Acta Orthopaedica et Traumatologica Turcica, vol. 58, no. 1, p. 1, 2024

    2024

  22. [30]

    Cut the crap: a critical response to “chatgpt is bullshit

    D. Gunkel and S. Coghlan, “Cut the crap: a critical response to “chatgpt is bullshit”,” Ethics and Information Technology , vol. 27, no. 2, p. 23, 2025

    2025

  23. [31]

    Chatgpt is bullshit,

    M. T. Hicks, J. Humphries, and J. Slater, “Chatgpt is bullshit,” Ethics and Information Technology , vol. 26, no. 2, pp. 1–10, 2024

    2024

  24. [32]

    Chatgpt and the educational ai chatter: Full of bullshit or trying to tell us something?

    E. Costello, “Chatgpt and the educational ai chatter: Full of bullshit or trying to tell us something?” Postdigital Science and Education , vol. 6, no. 2, pp. 425–430, 2024

    2024

  25. [33]

    Is chatgpt full of bullshit?

    L. Gorrieri, “Is chatgpt full of bullshit?” Journal of Ethics and Emerging Technologies , vol. 34, no. 1, pp. 1–16, 2024

    2024

  26. [34]

    On bullshit, large language models, and the need to curb your enthusiasm,

    D. W. Tigard, “On bullshit, large language models, and the need to curb your enthusiasm,” AI and Ethics , pp. 1–11, 2025

    2025

  27. [35]

    Beware of botshit: How to manage the epistemic risks of generative chatbots,

    T. R. Hannigan, I. P. McCarthy, and A. Spicer, “Beware of botshit: How to manage the epistemic risks of generative chatbots,” Business Horizons , vol. 67, no. 5, pp. 471–486, 2024

    2024

  28. [36]

    Tembine, M

    H. Tembine, M. Bouare, M. Dembele, A. Diallo, B. Diallo, A. Diarra, B. Doumbia, N. Molinier, A. Sidibe, A. Tapo, and S. Danioko, Machine Intelligence in Africa in 20 Questions . Bamako, Mali: Sawa Editions, 2023, published in collaboration with the National Library of Mali

    2023

  29. [37]

    Bas ¸ar, B

    T. Bas ¸ar, B. Djehiche, and H. Tembine, Mean-Field-Type Game Theory: Foundations and New Directions , 2025, vol. 1, forthcoming

    2025

  30. [38]

    2, forthcoming

    ——, Mean-Field-Type Game Theory: Applications , 2025, vol. 2, forthcoming

    2025

  31. [39]

    Machine intelligence in africa: a survey,

    A. A. Tapo, A. Traore, S. Danioko, and H. Tembine, “Machine intelligence in africa: a survey,” DSAI, 2024. [Online]. Available: https://arxiv.org/abs/2402.02218

    2024 arXiv

  32. [40]

    Mean-field-type transformers,

    H. Tembine, M. A. Khan, and I. Bamia, “Mean-field-type transformers,” Mathematics, vol. 12, no. 22, p. 3506, 2024. ACKNOWLEDGMENTS Authors gratefully acknowledge support from TIMADIE, Guinaga, Grabal, LnG Lab for the MFTG for Machine Intelligence project. Authors gratefully ac...

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.