Pith. sign in

REVIEW 2 major objections 5 minor 50 references

Phantom Evidence: How and Why Generative AI Manufactures False Positives in Science

T0 review · 2 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Generative AI erodes scientific evidence by inflating the chance that a convincing output appears even when the target is absent; the resulting likelihood ratio collapses toward one, so polish adds no evidence and the true-positive fraction

desk verdict A candid, self-aware perspective that gives a useful name and ledger to AI-inflated evidence, but one of its three headline claims — that polishing adds no evidence — overreaches what the data-processing inequality can prove. read the letter →

arxiv 2607.25991 v1 pith:Z7GQG3GY submitted 2026-07-28 q-bio.NC

Yukiyasu Kamitani , Ken Shirakawa This is my paper
classification q-bio.NC
keywords phantomevidencelikelihoodratiogenerativeAInegativecontrolsfalsepositivesselectionbiasdataleakagereproducibility
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 argues that when generative AI makes persuasive output cheap, a result that 'looks convincing' stops functioning as evidence for its truth. Evidence is formalized as a likelihood ratio, and the ratio is shown to be controlled by the denominator: the probability of seeing a convincing output even when the target is absent. Because generative systems are deliberately tuned to produce plausible output, that denominator swells and drives the likelihood ratio toward one. Three consequences follow: higher resolution and fluency add no evidence, a single result's evidence has a ceiling that self-grading cannot exceed, and the fraction of published findings that are true falls back to the field's base rate. The proposed cure, framed as Bacon's table of absence restated in probability, is to measure the denominator with negative controls.

What carries the argument

The central object is the likelihood ratio Λ = Pr(C|T)/Pr(C|¬T) for the paired propositions 'the artifact achieves its target' (T) and 'the artifact is judged convincing' (C). The load-bearing identity is the one-way chain T → primary data → exhibit → C, which, through the data-processing inequality, caps the evidence any downstream processing can carry at log k_eff, where k_eff = 1/γ is the effective number of output options the system can actually reach. Phantom evidence, Δ_total = log(N·Pr(C|¬T)), is the operational quantity measuring how far the perceived weight of evidence exceeds the actual one, and it decomposes additively into the nominal-vs-effective confusion log(N/k_eff), selectio

What would settle it

Run a full reconstruction pipeline (generative prior plus classification plus the same selection rule) on sham data where the target is absent, such as time-shuffled or other-subject brain signals, and measure the fraction of outputs judged convincing and correctly identified among many same-category distractors. The paper predicts this rate stays near 1/N; observing it near the genuine-reconstruction accuracy would refute the mechanism.

Watch

Extended reading notes

Core claim

The central claim is that evidential value lives in Pr(C|¬T), the probability that a generative system produces an output judged convincing even when the target is absent. The paper defines phantom evidence as the log gap between the observer's assumed denominator, 1/N, and the actual denominator, Pr(C|¬T), formalized as Δ_total = log(N·Pr(C|¬T)), which decomposes into nominal-vs-effective confusion, selection multiplicity, and leakage. As this denominator approaches the numerator, the likelihood ratio Λ = Pr(C|T)/Pr(C|¬T) collapses toward one, so observing a convincing artifact carries no evidence regardless of how many positive cases are accumulated. The posterior probability of truth then

Load-bearing premise

The result holds only while the exhibit is a function of the primary data alone and the judgment C is a function of the exhibit alone; if information about the target enters from outside the primary data—through the generative model's prior or training set—the ceiling on evidence can be breached.

Editorial extensions

If this is right

  • Higher output resolution and greater fluency inflate the nominal space N without raising the effective number of options k_eff, so they add no evidence and only enlarge phantom evidence.
  • A single result's evidence is bounded above by log k_eff; polishing outputs or letting the generator grade itself cannot exceed this ceiling.
  • When denominator collapse sets in, the fraction of published findings that are true falls back to the base rate π, regardless of statistical power or the prior.
  • Preregistration, adversarial negative controls, and contamination control close off the selection and leakage routes, while factorized or compositional output designs can genuinely raise the ceiling.
  • Reporting formats that separate generation quality from evidence, using discriminative tests against well-chosen distractors, are needed to keep convincing outputs from being counted as evidence.

Reading between the lines

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

  • The argument is domain-neutral: any technology that cheaply produces persuasive artifacts, from automated experimental pipelines to synthetic-data generators, should be expected to induce the same denominator inflation and thus the same collapse in evidential value.
  • If the paper is right, journals and review procedures should require the reporting of selection multiplicity and negative-control denominators, and treat LLM-as-judge evaluations with suspicion unless the judging model is outside the optimization loop.
  • A testable extension is to measure k_obs_eff = 1/Pr(C|¬T) for a field's full evaluation pipeline and compare it with that field's observed replication rate; the framework predicts a direct relationship between denominator inflation and reproducibility failures.
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

2 major / 5 minor

Summary. The paper argues that generative AI inflates the probability that an artifact looks convincing even when the claimed target is absent, and calls the observer's overestimate "phantom evidence." It defines target T, convincing judgment C, and evidence as log likelihood ratio Λ = Pr(C|T)/Pr(C|¬T), then derives Δ = log(N/k_eff) with k_eff = 1/Pr(C|¬T), extending the expression to selection multiplicity m_eff and leakage λ. It claims that polishing cannot add evidence, that a single result's evidence has ceiling log k_eff, that PPV collapses to the base rate as Λ→1, and it prescribes negative controls and output diversity. The algebraic identities in Eqs (1)-(10) are correct given the stated definitions.

Significance. The paper condenses an important and timely concern into a compact formal quantity and gives actionable prescriptions, especially its emphasis on measuring Pr(C|¬T) with negative controls. It is unusually transparent about scope: Section 8 discloses that the worked example uses the authors' own prior work, that the decomposition is not a point estimate, and that Direction 1 is an empirical bet. The algebra is clear and checkable. However, one of the three headline consequences rests on a confusion between expected mutual information and pointwise weight of evidence. That gap is load-bearing for the abstract's central claim, so the manuscript needs revision before the advertised conclusions can be accepted as stated.

major comments (2)
  1. [Section 2, Eq. (3), and Footnote 1] The inference from the data-processing inequality to "polishing cannot add evidence" is not valid as stated. Eq. (3) bounds expected mutual information I(T;C) ≤ I(T;exhibit); it does not bound the pointwise weight of evidence log Λ = log[Pr(C|T)/Pr(C|¬T)] defined in Eq. (2). A downstream function can increase Λ(C) for the particular observed C by pooling the observed exhibit with other states, so DPI does not imply that polishing leaves Λ unchanged or cannot raise it for a realized output. Footnote 1 and Appendix C acknowledge that I and log Λ are different, but the prose after Eq. (3) and the abstract assert the pointwise consequence without this qualification. This is load-bearing because the abstract's first consequence is a central result. Please restrict the no-added-evidence claim to expected information, or prove the pointwise form under additional assumptions; the ceiling log Λ≤l
  2. [Section 3 and Section 8] The abstract states "Higher resolution and greater fluency add no evidence" as a general law, but this is not derived as a theorem. It holds only when added resolution is filled by the generative prior without increasing the effective reachable set (output dimension collapse) and under the one-way-chain assumption. Section 7, Direction 1, lists six conditions under which k_eff can instead grow, and Section 8 concedes that the framework weakens in regions where independent verification is cheap and where numerator saturation fails. The headline claim should therefore be stated conditionally, and the abstract should carry the same qualifiers as the body.
minor comments (5)
  1. [Abstract and Section 2] The phrase "three things follow" should include the qualifier "under the one-way-chain assumption and output dimension collapse" to avoid overstating the scope.
  2. [Section 2, Footnote 1] The distinction between expected mutual information I(T;·) and pointwise log Λ is placed in a footnote while the main text uses "information" in both senses. The caveat should be moved into the prose so the advertised conclusion is not read as a pointwise statement.
  3. [Section 6] The statement Pr(C|¬T)≈1/N=0.01 for forced-choice identification assumes well-designed distractors. This is qualified later in Section 7(4), but the earlier sentence could mislead if read in isolation.
  4. [Section 8] The disclosure that the worked example derives from the authors' own prior work and was selected from candidate cases is commendable; consider also placing a one-sentence version at the start of Section 6 so readers encounter the disclosure before the numbers.
  5. [Appendix A] The relationship between the Gaussian extreme-value threshold t* ~ sqrt(2 ln m) and the Bernoulli selection penalty log m_eff is not fully spelled out. A short bridge equation showing that the reported event "a best path exceeds t" has selection penalty ≈ log m under the model of Eq. (6) would help.

Circularity Check

2 steps flagged · score 4.0 of 10

Core identities are definitional bookkeeping, but the empirical mechanism and disclosed self-citations do not force the central claim.

  1. self definitional [Section 3, Eq. (5); Section 5, Eq. (10)]
    "we call its gap from Λperceived phantom evidence: ∆ = log Λperceived − log Λactual = log(Nγ) = log N/keff ... logit PPVperceived − logit PPVactual = log Λperceived/Λactual = log(NPr(C|¬T)) = ∆total"

    ∆ is introduced in Eq. (5) as exactly log(Λperceived/Λactual), the log gap between believed and actual likelihood ratios. Eq. (10) then 'derives' that the log-odds PPV overestimate equals ∆total. Since logit PPV = log prior odds + log Λ, the difference between perceived and actual logit PPV is log(Λperceived/Λactual) by algebra. The proposition is therefore a restatement of the definition, not an independent result; it is load-bearing for Section 5's claim that only N/keff, meff, and leakage inflate the gap, which are the terms already built into the definition.

  2. self definitional [Section 3, Eq. (4); Section 7, Direction 1]
    "keff := 1/γ ... The actual likelihood ratio is then Λactual = Pr(C|T)/γ = Pr(C|T)·keff ... the ceiling on the effective evidence, log Λactual ≤ log keff, rises toward log N"

    The 'ceiling' is obtained by substituting keff := 1/γ into Λactual = Pr(C|T)/γ and using Pr(C|T) ≤ 1. It is true by construction of keff, not a bound discovered from outside the definitions. The paper presents it as the binding constraint ('the bound log Λactual ≤ log keff is the binding constraint'), but it is an algebraic consequence of the operational definition of keff.

full rationale

The central quantity ∆ is defined rather than empirically derived: Eq. (5) defines phantom evidence as the log gap between perceived and actual likelihood ratios, and Eq. (10) is the same gap rewritten in logit-PPV form. Similarly, the ceiling log Λactual ≤ log keff follows immediately from keff := 1/γ and Pr(C|T) ≤ 1. These are algebraic restatements, not discoveries from independent premises. However, the paper is transparent about this: Section 8 states that the framework is 'Bayesian inference itself and in that sense mathematically not new.' The substantive empirical mechanism—that generative AI inflates Pr(C|¬T)—is an external claim supported by prior literature and by the Section 6 worked example. The self-citations (refs 7, 28, 29) are disclosed, including the statement that 'the worked example (Section 6) derives from the authors' own prior work and was selected from among other candidate cases'; they support the example, not the derivation, and are not used as a load-bearing authority. The DPI argument in Section 2 does contain a genuine internal gap: the footnote admits that I is expected information and is 'a different quantity from the weight of evidence log Λ carried by a particular observation,' yet the prose and abstract apply the DPI bound to conclude that polishing 'cannot add a single bit' and 'neither adds evidence.' That is a correctness/validity concern about applying an expected-information bound to a pointwise quantity, but it is not a reduction of a result to its own input, so it is not counted as circularity here. Overall, the paper's headline identities are definitional bookkeeping, but the empirical core is not forced by a self-citation chain, yielding a moderate circularity score.

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

The framework rests on standard probability and information theorems (Bayes, DPI) plus several stated modeling choices that are not independently verified: the one-way chain, the independence approximation in Eq. (6), numerator saturation, and the observer's default 1/N. No parameters are fitted to data; all listed values are illustrative.

free parameters (6)
  • Nominal space N = 10^6 (running example); 100 (worked example identification)
    The size of the candidate space the observer assumes when computing 1/N surprise. Chosen for illustration, not fitted.
  • Effective options k_eff = 50 (running example); ≈1.11 (realism criterion, since Pr(C|¬T)=0.90); 100 (identification criterion, since Pr(C|¬T)=0.01
    Defined as 1/γ. Values are illustrative and taken from the worked example.
  • Selection multiplicity m_eff = 20 (running example)
    Effective number of independent analyses. The paper acknowledges it is not uniquely determined (Section 8).
  • Power q = Pr(C|T) = 0.9 (running example); 0.95 (worked example realism); 0.9 (identification)
    Probability a genuine artifact looks convincing. Used in illustrative calculations; not fitted to data.
  • Leakage degree λ = 1 (running example, so the leakage term vanishes); 0–1 in general
    Degree to which evaluation data contaminates the pipeline. A modeling parameter in Eq. (6).
  • Leakage positive probability r = not specified
    Probability that a leaked path yields a positive even when the target is absent. Appears in Eq. (6) but is not assigned a value.
assumptions (6)
  • standard math Bayes' theorem and the likelihood ratio as the measure of evidential value (Eqs 1–2)
    Cited to Good and Royall; standard probability theory.
  • standard math Data-processing inequality applies to the chain T → primary data → exhibit → C (Eq. 3)
    The inequalities H(T)≥I(T;primary data)≥I(T;exhibit)≥I(T;C) follow from the stated Markov chain; standard information theory.
  • domain assumption One-way chain: exhibit ⊥ T | primary data and C ⊥ T | exhibit
    Stated in Section 2 as the condition for the DPI argument; load-bearing for 'no evidence from polishing.'
  • ad hoc to paper Effective denominator combines independent analyses and leakage as Pr(C|¬T)≈1−(1−γλ)^{m_eff} (Eq. 6)
    Introduced in Section 4 as an 'independence-and-leakage approximation'; not derived from first principles.
  • domain assumption Numerator saturation Pr(C|T)≈1
    Assumed in Section 2; acknowledged as an approximation in Section 8.
  • domain assumption The observer's naive denominator is 1/N
    Used to define Λ_perceived (Section 3); an empirical/psychological assumption about how audiences count surprise.
invented entities (2)
  • Phantom evidence (Δ_total) independent evidence
    purpose: Quantifies the gap between perceived and actual evidence from a convincing output, combining nominal/effective confusion, selection, and leakage.
    Defined in Eqs (5) and (7); the paper gives a measurement procedure (negative controls and k_obs_eff), so it is falsifiable in principle. But it is a definitional reconstruction of known Bayes-factor components.
  • Idola Machinae
    purpose: Rhetorical label for the industrialization of Bacon's Idols of the Marketplace and Theatre through generative AI.
    Introduced in the conclusion as a name, not a measurable quantity; no independent empirical content.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Phantom Evidence: How and Why Generative AI Manufactures False Positives in Science." pith.science (2026). https://pith.science/paper/Z7GQG3GY

@misc{pith2026260725991,
  author       = {Pith},
  title        = {Pith review of: Phantom Evidence: How and Why Generative AI Manufactures False Positives in Science},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z7GQG3GY}},
  note         = {Machine review of arXiv:2607.25991}
}
read the original abstract

Four centuries ago Francis Bacon warned against the anticipations of nature, hasty generalization that wins assent on a few facts, and set against it the table of absence: checking that a property fails to appear where it should not. The demand was that looking convincing should not, on its own, count as evidence. Science has professed that demand ever since, while in practice letting persuasiveness do the work of evidence. It could be let to do so because making something persuasive was itself hard. Generative AI removes that difficulty, and an old error returns on a scale and at a speed it never had before. We locate the problem not in evidence growing weaker but in how surprise is counted. An observer marvels at a convincing output as a single point hit among a vast range of possibilities, yet what a system can actually reach is a small part of that range. The gap between the breadth imagined and the narrowness actually reached is what we call phantom evidence, and we formalize it as one quantity that also absorbs the trial and error and the data leakage a research process adds. Three things follow. Higher resolution and greater fluency add no evidence. The evidence a single result can carry has a ceiling that neither polishing the output nor letting a generative system grade itself can exceed. And the fraction of published findings that are true falls back to what it was before anything was observed. The prescription lies in the same place: genuinely widen what a system can reach, and measure whether convincing outputs still appear when the target is absent -- Bacon's table of absence, restated in the language of probability. In a world where the persuasive has become cheap, the credibility of science rests not on more convincing outputs but on procedures that show they could not have arisen by chance.

Figures

Figures reproduced from arXiv: 2607.25991 by the authors.

Figure 1
Figure 1. One mechanism, three faces. Clever Hans, the overestimation of large language model [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The intuition of phantom evidence: nominal [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. The three-term decomposition of phantom evidence (running example, in bits [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The ladder of evidence, the two ceilings, and the two directions of the prescription [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Shallow against whole (self-contained) negative control. Top: a shallow check that [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: The lifetime of evidence: two clocks. During the research phase ( [PITH_FULL_IMAGE:figures/full_fig_p023_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 5 linked inside Pith

  1. [1]

    A.et al.Comparing scientific abstracts generated by ChatGPT to real abstracts with detectors and blinded human reviewers.npj Digital Medicine6,75 (2023)

    Gao, C. A.et al.Comparing scientific abstracts generated by ChatGPT to real abstracts with detectors and blinded human reviewers.npj Digital Medicine6,75 (2023)

    2023

  2. [2]

    & Crockett, M

    Messeri, L. & Crockett, M. J. Artificial intelligence and illusions of understanding in sci- entific research.Nature627,49–58 (2024)

    2024

  3. [3]

    von Osten): A Contribution to Experimental Animal and Human Psychology(Henry Holt and Company, New York, 1911)

    Pfungst, O.Clever Hans (the Horse of Mr. von Osten): A Contribution to Experimental Animal and Human Psychology(Henry Holt and Company, New York, 1911)

    1911

  4. [4]

    Lapuschkin, S.et al.Unmasking Clever Hans predictors and assessing what machines really learn.Nature Communications10,1096 (2019)

    2019

  5. [5]

    Geirhos, R.et al.Shortcut Learning in Deep Neural Networks.Nature Machine Intelligence 2,665–673 (2020)

    2020

  6. [6]

    & Koyejo, S.Are Emergent Abilities of Large Language Models a Mirage?inAdvances in Neural Information Processing Systems 36 (NeurIPS 2023)(2023)

    Schaeffer, R., Miranda, B. & Koyejo, S.Are Emergent Abilities of Large Language Models a Mirage?inAdvances in Neural Information Processing Systems 36 (NeurIPS 2023)(2023)

    2023

  7. [7]

    Shirakawa, K.et al.Spurious reconstruction from brain activity.Neural Networks190, 107515 (2025)

    2025

  8. [8]

    & Silverthorne, M.) (Cambridge University Press, Cambridge, 2000)

    Bacon, F.The New Organon(eds Jardine, L. & Silverthorne, M.) (Cambridge University Press, Cambridge, 2000)

    2000

Show all 50 references
  1. [9]

    J.Probability and the Weighing of Evidence(Charles Griffin, London, 1950)

    Good, I. J.Probability and the Weighing of Evidence(Charles Griffin, London, 1950)

    1950

  2. [10]

    Royall, R.Statistical Evidence: A Likelihood Paradigm(Chapman and Hall/CRC, London, 1997)

    1997

  3. [11]

    & Leibler, R

    Kullback, S. & Leibler, R. A. On information and sufficiency.The Annals of Mathematical Statistics22,79–86 (1951)

    1951

  4. [12]

    G.Statistical Inference as Severe Testing: How to Get Beyond the Statistics Wars (Cambridge University Press, Cambridge, 2018)

    Mayo, D. G.Statistical Inference as Severe Testing: How to Get Beyond the Statistics Wars (Cambridge University Press, Cambridge, 2018)

    2018

  5. [13]

    & Zhang, Y.Do GANs learn the distribution? Some theory and empiricsinInternational Conference on Learning Representations (ICLR)(2018)

    Arora, S., Risteski, A. & Zhang, Y.Do GANs learn the distribution? Some theory and empiricsinInternational Conference on Learning Representations (ICLR)(2018)

    2018

  6. [14]

    Hill, M. O. Diversity and evenness: A unifying notation and its consequences.Ecology54, 427–432 (1973)

    1973

  7. [15]

    & Aila, T.Improved precision and recall metric for assessing generative modelsinAdvances in Neural Information Processing Systems 32 (NeurIPS)(2019)

    Kynkäänniemi, T., Karras, T., Laine, S., Lehtinen, J. & Aila, T.Improved precision and recall metric for assessing generative modelsinAdvances in Neural Information Processing Systems 32 (NeurIPS)(2019). arXiv:1904.06991 [stat.ML]

    2019 arXiv

  8. [16]

    P., Nelson, L

    Simmons, J. P., Nelson, L. D. & Simonsohn, U. False-positive psychology: Undisclosed flexi- bility in data collection and analysis allows presenting anything as significant.Psychological Science22,1359–1366 (2011)

    2011

  9. [17]

    & Loken, E

    Gelman, A. & Loken, E. The Statistical Crisis in Science.American Scientist102,460–465 (2014). 24

    2014

  10. [18]

    Dwork,C.et al.Thereusableholdout:Preservingvalidityinadaptivedataanalysis.Science 349,636–638 (2015)

    2015

  11. [19]

    K., Bellgowan, P

    Kriegeskorte, N., Simmons, W. K., Bellgowan, P. S. F. & Baker, C. I. Circular analysis in systems neuroscience: the dangers of double dipping.Nature Neuroscience12,535–540 (2009)

    2009

  12. [20]

    & Hardt, M.The Ladder: A reliable leaderboard for machine learning competitions arXiv:1502.04585 [cs.LG]

    Blum, A. & Hardt, M.The Ladder: A reliable leaderboard for machine learning competitions arXiv:1502.04585 [cs.LG]

    arXiv

  13. [21]

    & Shankar, V.Do ImageNet classifiers generalize to ImageNet?arXiv:1902.10811 [cs.CV]

    Recht, B., Roelofs, R., Schmidt, L. & Shankar, V.Do ImageNet classifiers generalize to ImageNet?arXiv:1902.10811 [cs.CV]

    1902 arXiv

  14. [22]

    Ioannidis, J. P. A. Why most published research findings are false.PLoS Medicine2,e124 (2005)

    2005

  15. [23]

    Meehl, P. E. Theoretical risks and tabular asterisks: Sir Karl, Sir Ronald, and the slow progress of soft psychology.Journal of Consulting and Clinical Psychology46,806–834 (1978)

    1978

  16. [24]

    Estimating the reproducibility of psychological science.Sci- ence349,aac4716 (2015)

    Open Science Collaboration. Estimating the reproducibility of psychological science.Sci- ence349,aac4716 (2015)

    2015

  17. [25]

    Lipton, Z. C. & Steinhardt, J. Troubling trends in machine learning scholarship.Queue 17,45–77 (2019)

    2019

  18. [26]

    & Narayanan, A

    Kapoor, S. & Narayanan, A. Leakage and the reproducibility crisis in machine-learning- based science.Patterns4,100804 (2023)

    2023

  19. [27]

    & Shirakawa, K

    Kamitani, Y., Tanaka, M. & Shirakawa, K. Visual image reconstruction from brain activity via latent representation.Annual Review of Vision Science11,611–634 (2025)

    2025

  20. [28]

    & Kamitani, Y

    Otsuka, K., Nagano, Y. & Kamitani, Y. Overcoming output dimension collapse: When sparsity enables zero-shot brain-to-image reconstruction at small data scales.Transactions on Machine Learning Research.arXiv:2509.15832 (2026)

    2026 arXiv

  21. [29]

    Shirakawa, K., Nagano, Y., Tanaka, M., Cheng, F. L. & Kamitani, Y.Advancing credibility and transparency in brain-to-image reconstruction research: Reanalysis of Koide-Majima, Nishimoto, and Majima (Neural Networks, 2024)arXiv:2511.07960 [q-bio.NC]

    2024

  22. [30]

    & Majima, K

    Koide-Majima, N., Nishimoto, S. & Majima, K. Mental image reconstruction from human brain activity: Neural decoding of mental imagery via deep neural network-based Bayesian estimation.Neural Networks170,349–363 (2024)

    2024

  23. [31]

    A., Ebersole, C

    Nosek, B. A., Ebersole, C. R., DeHaven, A. C. & Mellor, D. T. The preregistration revo- lution.Proceedings of the National Academy of Sciences115,2600–2606 (2018)

    2018

  24. [32]

    N., Naselaris, T., Prenger, R

    Kay, K. N., Naselaris, T., Prenger, R. J. & Gallant, J. L. Identifying natural images from human brain activity.Nature452,352–355 (2008)

    2008

  25. [33]

    Job market signaling.The Quarterly Journal of Economics87,355–374 (1973)

    Spence, M. Job market signaling.The Quarterly Journal of Economics87,355–374 (1973)

    1973

  26. [34]

    Judging LLM-as-a-Judge with MT-Bench and Chatbot ArenainAdvances in Neural Information Processing Systems36(2023), 46595–46623

    Zheng, L.et al. Judging LLM-as-a-Judge with MT-Bench and Chatbot ArenainAdvances in Neural Information Processing Systems36(2023), 46595–46623

    2023

  27. [35]

    Panickssery, A., Bowman, S. R. & Feng, S.LLM evaluators recognize and favor their own generationsinAdvances in Neural Information Processing Systems37(2024),68772–68802

    2024

  28. [36]

    S., Kumar, A., Balasubramanian, S., Wang, W

    Sadasivan, V. S., Kumar, A., Balasubramanian, S., Wang, W. & Feizi, S.Can AI-generated text be reliably detected?arXiv:2303.11156 [cs.CL]

    arXiv

  29. [37]

    Goodfellow, I.et al.Generative adversarial networks.Communications of the ACM63, 139–144 (2020)

    2020

  30. [38]

    MacKay, D. J. C.Information Theory, Inference, and Learning Algorithms(Cambridge University Press, Cambridge, 2003). 25

    2003

  31. [39]

    Large-scale simultaneous hypothesis testing: The choice of a null hypothesis

    Efron, B. Large-scale simultaneous hypothesis testing: The choice of a null hypothesis. Journal of the American Statistical Association99,96–104 (2004)

    2004

  32. [40]

    Eklund, A., Nichols, T. E. & Knutsson, H. Cluster failure: Why fMRI inferences for spatial extent have inflated false-positive rates.Proceedings of the National Academy of Sciences 113,7900–7905 (2016)

    2016

  33. [41]

    Pocock,S.J.Groupsequentialmethodsinthedesignandanalysisofclinicaltrials.Biometrika 64,191–199 (1977)

    1977

  34. [42]

    O’Brien, P. C. & Fleming, T. R. A Multiple Testing Procedure for Clinical Trials.Biomet- rics35,549–556 (1979)

    1979

  35. [43]

    Lan, K. K. G. & DeMets, D. L. Discrete sequential boundaries for clinical trials.Biometrika 70,659–663 (1983)

    1983

  36. [44]

    & Walsh, D

    Johari, R., Koomen, P., Pekelis, L. & Walsh, D. Always Valid Inference: Continuous Mon- itoring of A/B Tests.Operations Research70,1806–1821 (2022)

    2022

  37. [45]

    Unpublished results hide the decline effect.Nature470,437 (2011)

    Schooler, J. Unpublished results hide the decline effect.Nature470,437 (2011)

    2011

  38. [46]

    Poynard, T., Munteanu, M., Ratziu, V.,et al.Truth Survival in Clinical Research: An Evidence-Based Requiem?Annals of Internal Medicine136,888–895 (2002)

    2002

  39. [47]

    Ioannidis, J. P. A. Contradicted and Initially Stronger Effects in Highly Cited Clinical Research.JAMA294,218–228 (2005)

    2005

  40. [48]

    F.et al.Evaluating the replicability of social science experiments in Nature and Science between 2010 and 2015.Nature Human Behaviour2,637–644 (2018)

    Camerer, C. F.et al.Evaluating the replicability of social science experiments in Nature and Science between 2010 and 2015.Nature Human Behaviour2,637–644 (2018)

    2010

  41. [49]

    Negative results are disappearing from most disciplines and countries.Sciento- metrics90,891–904 (2011)

    Fanelli, D. Negative results are disappearing from most disciplines and countries.Sciento- metrics90,891–904 (2011)

    2011

  42. [50]

    Little, J. D. C. A Proof for the Queuing Formula:L=λW.Operations Research9,383–387 (1961). 26

    1961

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.