Pith. sign in

REVIEW 3 major objections 5 minor 28 references

AI-generated fake news is easier to detect than human-written fake news

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 06:10 UTC pith:PLABAI5M

load-bearing objection A plausible and relevant empirical question, but the manuscript as written is unverifiable—missing corpus details, an impossible p-value, and mismatched captions—so it needs major revision and full data release before it can be trusted. the 3 major comments →

arxiv 2607.21967 v1 pith:PLABAI5M submitted 2026-07-24 stat.AP

When Machines Lie Differently: Detecting AI vs Human Fake News

Canliv Ibenye , Aya-Vera Jimenez , Samuel Jaeger , Dhrubajyoti Ghosh This is my paper
classification stat.AP
keywords fake news detectionAI-generated textmisinformationtext classificationreadability analysissentiment analysismachine learningAUC
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks whether fake news becomes easier or harder to detect when it is written by a large language model instead of a person. Using the same feature representation and the same set of classifiers for both tasks, it finds that real-versus-AI-fake classification achieves near-perfect AUC (above 0.99 for most models) while real-versus-human-fake classification scores between 0.89 and 0.96. Because the modeling pipeline is identical, the authors argue this gap reflects intrinsic statistical differences in the text, not the choice of algorithm. The result matters because deployed detection systems face both kinds of misinformation, and the asymmetry suggests a single unified detector may be insufficient.

Core claim

Under a common feature representation built from lexical diversity, readability indices, and emotion lexicons, AI-generated fake news is substantially more separable from real news than human-written fake news. Across all six models tested, the AUC for the real-versus-AI task is uniformly higher than for the real-versus-human task, with an average improvement of about 0.049 (paired t-test t = 8.67, p < 0.001; Wilcoxon signed-rank p = 0.016). Feature-level analysis shows AI-generated fake news has more uniform readability and emotional patterns, producing less overlap with real news, while human-written fake news mimics the heterogeneity of legitimate journalism. The authors conclude that the

What carries the argument

The central object is the feature mapping φ(x) that concatenates type-token ratio, Flesch and Flesch-Kincaid readability scores, SMOG and Coleman-Liau indices, and NRC emotion-lexicon features (anger, fear, joy, sadness, trust, anticipation, disgust, surprise, plus positive/negative sentiment). The argument is carried by comparing two binary classification tasks — real vs human fake and real vs AI fake — under an identical pipeline of feature extraction, preprocessing, model classes, and evaluation, so any systematic AUC difference is attributed to the distribution of the fake class.

Load-bearing premise

The comparison assumes that one unspecified language model and one prompt scheme adequately represent 'AI-generated fake news', and that the real-news corpus is identical and topically matched across both tasks; otherwise the AUC gap could reflect topic drift or dataset artifacts rather than the source of the fake content.

What would settle it

Re-run the same pipeline with several current LLMs (e.g., different architectures and parameter scales) and multiple prompt templates, keeping the real-news corpus fixed and topic-matched, and check whether any AI-generated fake news set achieves AUC comparable to the 0.89–0.96 range seen for human fake news. If such a set exists, the claim that AI-generated fake news is intrinsically more separable would collapse.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Current detection systems may be well-equipped to identify AI-generated fake news in its present form, but they remain substantially less reliable for sophisticated human-written misinformation.
  • Because the performance gap persists across linear, kernel, tree-based, neural, and ensemble models, the separability appears to be a property of the feature space itself rather than any particular classifier.
  • Ensembling models yields the highest AUC in both tasks but does not change the ordering, reinforcing that the data-generating process drives the result.
  • If generative models become more diverse in readability and emotional expression, the gap may narrow, requiring detection methods to adapt as LLMs evolve.
  • Text-only surface features face inherent limits against human-authored deception, suggesting that contextual, social, and network signals may be necessary for that harder task.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's result is a snapshot tied to one unspecified LLM and one prompt scheme; using multiple current LLMs with varied prompts could show that some configurations produce AI fake news nearly as hard to detect as human fake news, which would temper the claim of intrinsic separability.
  • A testable extension would be to measure whether the gap shrinks when AI-generated text is post-edited by humans to mimic the readability and emotion variance of real news; if it closes, the finding is about current decoding habits rather than a fundamental property.
  • The authors' feature set is deliberately compact and interpretable; one could check whether transformer-based embeddings erase or widen the gap, which would clarify how much of the effect lives at the surface-statistics level.
  • The asymmetry implies that platform moderation policies treating all fake news alike may be misaligned with detection reality, but the ethical and adversarial implications of that asymmetry are left for future work.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper compares two binary classification tasks: distinguishing real news from human-written fake news and from AI-generated fake news. Both tasks use the same feature set (lexical diversity, readability indices, and NRC emotion features), the same preprocessing, and six machine-learning models. The reported held-out AUCs are uniformly higher for the real-versus-AI-fake task (0.963–0.996) than for the real-versus-human-fake task (0.888–0.960). Because the modeling pipeline is identical, the authors attribute the gap to intrinsic distributional differences between human- and AI-generated fake text. The manuscript includes feature-level analyses and paired statistical tests (t-test and Wilcoxon signed-rank) across models.

Significance. If the empirical result is trustworthy, this is a timely and useful contribution: it quantifies an asymmetry in detectability under a controlled pipeline and with interpretable features. The strengths are the explicit attempt to hold the modeling pipeline fixed, the transparent feature set, and the authors' acknowledgment of some limitations. However, the central claim is not yet established because the two fake-news corpora are not described in enough detail to rule out corpus artifacts, and the statistical reporting lacks confidence intervals and a clear inference framework. The paper is also not reproducible in its current form, with no data or code availability.

major comments (3)
  1. [Section II (Method)] The design claim that 'the two learning problems differ only in the distribution of the fake class' is load-bearing but unsupported. The manuscript does not name the real-news dataset, the human-fake 'curated misinformation datasets', the LLM used, the prompting scheme, decoding parameters, or the topical matching procedure. If the human-fake corpus and AI-fake corpus differ in document length, genre, or topic coverage, the features used (TTR, readability, emotions) will encode those differences rather than source identity. Please provide dataset names, sample sizes, and a demonstration of comparability across the three classes, e.g., distributions of length and topic, or a matched-subset analysis.
  2. [Section III, Table I and statistical tests] The AUC values are given as single point estimates from one 80/20 split, without confidence intervals, standard errors, or repeated resampling. The paired t-test and Wilcoxon test are performed across M=6 models, treating the model set as paired observations; this does not support inference about document- or corpus-level variability. Also, the Wilcoxon direction is not specified: with six positive differences, the one-sided exact p is 1/64≈0.016, while the two-sided exact minimum is 0.031. The reported p=0.016 is therefore ambiguous. Please report bootstrap or repeated-CV confidence intervals for the AUC gap and describe the exact test implementation and direction.
  3. [Section IV (Discussion) and Section V (Conclusion)] The phrase 'intrinsic distributional properties of the text' overstates what a fixed, hand-selected feature set can establish. The gap is measured in one specific feature space (TTR, readability indices, NRC emotions), and the Discussion correctly limits the claim later. Please either soften the abstract and conclusion to 'under this feature representation', or add a robustness check with a complementary feature family (e.g., contextual embeddings or more stylometric features) to show that the direction of the gap does not depend on the feature choice.
minor comments (5)
  1. [Figure captions] The figure captions are mismatched with the text: Fig. 1 is described in the text as an AUC bar chart but captioned as a Coleman–Liau density plot; Fig. 2 is described as a Coleman–Liau density plot but captioned as feature-wise mean differences; Fig. 3 is described as a feature heatmap but captioned as ROC curves. The cross-references need to be fixed.
  2. [Table I] The table header still reads 'TABLE TYPE STYLES'; replace with a descriptive caption.
  3. [Reproducibility] No data availability statement or code repository is provided, which is a barrier to reproducing the reported AUCs.
  4. [Dataset description] Section II says 'balanced samples across classes' but gives no class counts; report the number of documents in each class.
  5. [Figure 4] The text refers to Figure 4 for the ROC curves, but no Fig. 4 caption appears in the manuscript.

Circularity Check

0 steps flagged

No circular reduction found; the AUC gap is an empirical comparison whose main weaknesses are corpus-matching and numerical reliability, not self-definition or load-bearing self-citation.

full rationale

The central comparison is an empirical measurement: Δ_m = AUC_m^(A) − AUC_m^(H) is computed on held-out test data after fitting models on an 80% training split using fixed feature extraction. No output quantity is defined in terms of the conclusion, and no fitted parameter is renamed as a prediction. The self-citations ([1], [2], [12], [21]) are contextual: [12] is cited to note prior work on AI-vs-human text, [21] for the generic idea of ensembling, and [1]/[2] for the societal impact of misinformation. None is the source of the data, features, or result, so the citation chain is not load-bearing. The real threats are validity threats, not circularity: Section II asserts 'the two learning problems differ only in the distribution of the fake class' without specifying the LLM, prompts, decoding parameters, or human-fake corpus; genre/length/topic confounds could therefore explain the gap. That is an uncontrolled-comparison problem, not a circular one. The reported Wilcoxon p=0.016 with M=6 is arithmetically impossible (minimum two-sided exact p ≈ 0.031), and figure/table captions are mismatched; these are correctness/reliability problems. Because the paper contains only minor non-load-bearing self-citations and no exhibited reduction of a result to its inputs, the circularity score is 2.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The paper's central claim depends on an unvalidated feature space and unstated data-generation choices. No fitted constants are reported, and the main empirical quantity (the AUC gap) is not externally validated. The statistical test reporting is internally inconsistent.

free parameters (1)
  • Feature set (type-token ratio, readability indices, NRC emotion features)
    The hand-chosen feature representation is never validated against alternatives; the central claim that the gap reflects intrinsic text properties depends on this specific representation.
axioms (4)
  • domain assumption The same real-news corpus D_R is used in both tasks and is topically matched to the AI-fake corpus.
    Section II defines D_RH = D_R ∪ D_H and D_RA = D_R ∪ D_A, implying a shared D_R, but no dataset details confirm this. If the real sets differ, the AUC gap could be an artifact.
  • domain assumption A single unspecified LLM and prompting protocol adequately represents 'AI-generated fake news'.
    Section II states only 'AI-generated fake news is produced using a large language model under controlled prompting conditions'; no model, version, prompts, temperature, or sampling settings are given.
  • standard math AUC estimates from the models can be treated as paired exchangeable observations for t/Wilcoxon tests.
    Section III reports paired tests across models, but models are evaluated on the same test set and AUCs are correlated; with n=6 the reported Wilcoxon p=0.016 is impossible, indicating the testing procedure is not correctly described.
  • domain assumption Lexical, readability, and emotion features capture the 'intrinsic statistical differences' relevant to fake-news detectability.
    Section IV treats feature-space separability as intrinsic text properties; other representations (embeddings, syntax, discourse) could invert the gap.

pith-pipeline@v1.3.0-alltime-deepseek · 9155 in / 9467 out tokens · 97531 ms · 2026-08-01T06:10:36.478973+00:00 · methodology

0 comments
read the original abstract

The rapid advancement of large language models has introduced AI-generated fake news alongside traditional human-written misinformation, raising questions about whether detectability depends on the source of deceptive content. This study examines that issue through two controlled binary classification tasks: distinguishing real news from human-written fake news and from AI-generated fake news. Each article is represented using features related to lexical diversity, readability, and emotional characteristics, and evaluated with several machine learning models, including logistic regression, random forests, support vector machines, gradient boosting, neural networks, and ensemble methods. Performance is measured using the area under the receiver operating characteristic curve (AUC). Across all models, AI-generated fake news is detected with near-perfect accuracy, while human-written fake news is substantially more difficult to distinguish from real news. Because both tasks use the same modeling pipeline, this performance gap reflects intrinsic statistical differences in the text rather than methodological variation. Feature-level analysis shows that AI-generated fake news exhibits more uniform readability and emotional patterns, producing less overlap with real news. These findings reveal a key asymmetry in misinformation detection: current methods may be highly effective at identifying AI-generated content but remain less reliable against sophisticated human-authored misinformation. Detection systems should therefore account for the source of misinformation and continue adapting as generative models evolve.

Figures

Figures reproduced from arXiv: 2607.21967 by Aya-Vera Jimenez, Canliv Ibenye, Dhrubajyoti Ghosh, Samuel Jaeger.

Figure 2
Figure 2. Figure 2: Feature-wise mean differences between fake and real news. Differences are larger for AI-generated fake news, indicating greater deviation from real news. Further evidence is provided by the feature-level analysis shown in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: illustrates this result by displaying the AUC values for each model across the two tasks. The separation between the two groups of bars is pronounced and consistent, with all models achieving higher performance on the AI-generated fake news task. Notably, the ensemble model achieves the highest AUC in both settings, suggesting that combining multiple learners yields a more stable and accurate estimate of t… view at source ↗
Figure 3
Figure 3. Figure 3: ROC curves for the random forest classifier. The AI-generated fake news curve dominates, showing better detection performance than for human￾written fake news. Finally, the ROC curves for the random forest classifier, shown in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

28 extracted references · 4 canonical work pages

  1. [1]

    THANOS: A Predictive Model of Electoral Campaigns Using Twitter Data and Opinion Polls,

    D. Ghosh, W. A. Boettcher, R. Johnston, and S. Lahiri, “THANOS: A Predictive Model of Electoral Campaigns Using Twitter Data and Opinion Polls,” Data Sci. Sci., vol. 4, no. 1, p. 2484180, Dec. 2025, doi: 10.1080/26941899.2025.2484180

    arXiv 2025

  2. [2]

    Machine Learning Based Bot Detection on X With Temporal and Semantic Feature Integration,

    D. Ghosh, W. Boettcher, R. Johnston, and S. Lahiri, “Machine Learning Based Bot Detection on X With Temporal and Semantic Feature Integration,” IEEE Trans. Comput. Soc. Syst., 2026, Accessed: Apr. 10, 2026. [Online]. Available: https://ieeexplore.ieee.org/abstract/document/11394813/

    arXiv 2026

  3. [3]

    The science of fake news,

    D. M. J. Lazer et al., “The science of fake news,” Science, vol. 359, no. 6380, pp. 1094–1096, Mar. 2018, doi: 10.1126/science.aao2998

    doi:10.1126/science.aao2998 2018

  4. [4]

    The spread of true and false news online,

    S. Vosoughi, D. Roy, and S. Aral, “The spread of true and false news online,” Science, vol. 359, no. 6380, pp. 1146–1151, Mar. 2018, doi: 10.1126/science.aap9559

    doi:10.1126/science.aap9559 2018

  5. [5]

    Fake News Detection on Social Media: A Data Mining Perspective,

    K. Shu, A. Sliva, S. Wang, J. Tang, and H. Liu, “Fake News Detection on Social Media: A Data Mining Perspective,” SIGKDD Explor Newsl, vol. 19, no. 1, pp. 22–36, Sep. 2017, doi: 10.1145/3137597.3137600

    arXiv 2017

  6. [6]

    Fake News Detection: A Deep Learning Approach,

    A. Thota, P. Tilak, S. Ahluwalia, and N. Lohia, “Fake News Detection: A Deep Learning Approach,” SMU Data Sci. Rev., vol. 1, no. 3, Aug. 2018, [Online]. Available: https://scholar.smu.edu/datasciencereview/vol1/iss3/10

    2018

  7. [7]

    Information credibility on twitter,

    C. Castillo, M. Mendoza, and B. Poblete, “Information credibility on twitter,” in Proceedings of the 20th international conference on World wide web, in WWW ’11. New York, NY, USA: Association for Computing Machinery, Mar. 2011, pp. 675–684. doi: 10.1145/1963405.1963500

    arXiv 2011

  8. [8]

    Prominent Features of Rumor Propagation in Online Social Media,

    S. Kwon, M. Cha, K. Jung, W. Chen, and WangYajun, “Prominent Features of Rumor Propagation in Online Social Media,” Int. Conf. Data Min., 10 2013, doi: 10.1109/ICDM.2013.61

    work page doi:10.1109/icdm.2013.61 2013

  9. [9]

    Attention is All you Need,

    A. Vaswani et al., “Attention is All you Need,” in Advances in Neural Information Processing Systems, Curran Associates, Inc., 2017. Accessed: Apr. 10, 2026. [Online]. Available: https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd 053c1c4a845aa-Abstract.html

    2017

  10. [10]

    GLTR: Statistical Detection and Visualization of Generated Text,

    S. Gehrmann, H. Strobelt, and A. Rush, “GLTR: Statistical Detection and Visualization of Generated Text,” in Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, M. R. Costa-jussà and E. Alfonseca, Eds., Florence, Italy: Association for Computational Linguistics, Jul. 2019, pp. 111–116. doi: 10.18...

    doi:10.18653/v1/p19-3019 2019

  11. [11]

    Automatic Detection of Generated Text is Easiest when Humans are Fooled,

    D. Ippolito, D. Duckworth, C. Callison-Burch, and D. Eck, “Automatic Detection of Generated Text is Easiest when Humans are Fooled,” in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, D. Jurafsky, J. Chai, N. Schluter, and J. Tetreault, Eds., Online: Association for Computational Linguistics, Jul. 2020, pp. 1808–18...

    doi:10.18653/v1/2020.acl-main.164 2020

  12. [12]

    Human vs. Machine Deception: Distinguishing AI-Generated and Human-Written Fake News Using Ensemble Learning,

    S. Jaeger, C. Ibeneye, A. Vera-Jimenez, and D. Ghosh, “Human vs. Machine Deception: Distinguishing AI-Generated and Human-Written Fake News Using Ensemble Learning,” Apr. 10, 2026, arXiv: arXiv:2604.09960. doi: 10.48550/arXiv.2604.09960

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.09960 2026

  13. [13]

    Can linguists distinguish between ChatGPT/AI and human writing?: A study of research ethics and academic publishing,

    J. E. Casal and M. Kessler, “Can linguists distinguish between ChatGPT/AI and human writing?: A study of research ethics and academic publishing,” Res. Methods Appl. Linguist., vol. 2, no. 3, p. 100068, 2023

    2023

  14. [14]

    From human writing to artificial intelligence generated text: examining the prospects and potential threats of ChatGPT in academic writing,

    I. Dergaa, K. Chamari, P. Zmijewski, and H. B. Saad, “From human writing to artificial intelligence generated text: examining the prospects and potential threats of ChatGPT in academic writing,” Biol. Sport, vol. 40, no. 2, pp. 615–622, 2023

    2023

  15. [15]

    Lazy, not biased: Susceptibility to partisan fake news is better explained by lack of reasoning than by motivated reasoning,

    G. Pennycook and D. G. Rand, “Lazy, not biased: Susceptibility to partisan fake news is better explained by lack of reasoning than by motivated reasoning,” Cognition, vol. 188, pp. 39–50, Jul. 2019, doi: 10.1016/j.cognition.2018.06.011

    work page doi:10.1016/j.cognition.2018.06.011 2019

  16. [16]

    Type/Token Ratios: what do they really tell us?,

    B. Richards, “Type/Token Ratios: what do they really tell us?,” J. Child Lang., vol. 14, no. 2, pp. 201–209, Jun. 1987, doi: 10.1017/S0305000900012885

    doi:10.1017/s0305000900012885 1987

  17. [17]

    Simplification of Flesch Reading Ease Formula,

    J. N. Farr, J. J. Jenkins, and D. G. Paterson, “Simplification of Flesch Reading Ease Formula,” J. Appl. Psychol., vol. 35, no. 5, pp. 333– 337, 1951, doi: 10.1037/h0062427

    work page doi:10.1037/h0062427 1951

  18. [18]

    Evaluating Text Complexity and Flesch-Kincaid Grade Level,

    M. Solnyshkina, R. Zamaletdinov, L. Gorodetskaya, and A. Gabitov, “Evaluating Text Complexity and Flesch-Kincaid Grade Level,” J. Soc. Stud. Educ. Res., vol. 8, no. 3, pp. 238–248, Nov. 2017

    2017

  19. [19]

    SMOG Grading-a New Readability Formula,

    G. H. Mc Laughlin, “SMOG Grading-a New Readability Formula,” J. Read., vol. 12, no. 8, pp. 639–646, 1969

    1969

  20. [20]

    A computer readability formula designed for machine scoring,

    M. Coleman and T. L. Liau, “A computer readability formula designed for machine scoring,” J. Appl. Psychol., vol. 60, no. 2, pp. 283–284, 1975, doi: 10.1037/h0076540

    doi:10.1037/h0076540 1975

  21. [21]

    Ensemble survival analysis for preclinical cognitive decline prediction in Alzheimer’s disease using longitudinal biomarkers,

    D. Ghosh, S. Pal, M. Lutz, and S. Luo, “Ensemble survival analysis for preclinical cognitive decline prediction in Alzheimer’s disease using longitudinal biomarkers,” J. Alzheimer’s Dis., vol. 107, no. 3, pp. 1256–1266, Oct. 2025, doi: 10.1177/13872877251365621

    work page doi:10.1177/13872877251365621 2025

  22. [22]

    The Curious Case of Neural Text Degeneration,

    A. Holtzman, J. Buys, L. Du, M. Forbes, and Y. Choi, “The Curious Case of Neural Text Degeneration,” Feb. 14, 2020, arXiv: arXiv:1904.09751. doi: 10.48550/arXiv.1904.09751

    Pith/arXiv arXiv doi:10.48550/arxiv.1904.09751 2020

  23. [23]

    Typical decoding for natural language generation,

    C. Meister, T. Pimentel, G. Wiher, and R. Cotterell, “Typical decoding for natural language generation,” ArXiv Prepr. ArXiv220200666, vol. 4, 2022

    2022

  24. [24]

    Defending against neural fake news,

    R. Zellers et al., “Defending against neural fake news,” Adv. Neural Inf. Process. Syst., vol. 32, 2019, Accessed: Apr. 11, 2026. [Online]. Available: https://proceedings.neurips.cc/paper/2019/hash/3e9f0fc9b2f89e043bc 6233994dfcf76-Abstract.html

    2019

  25. [25]

    Computational methods in authorship attribution,

    M. Koppel, J. Schler, and S. Argamon, “Computational methods in authorship attribution,” J. Am. Soc. Inf. Sci. Technol., vol. 60, no. 1, pp. 9–26, Jan. 2009, doi: 10.1002/asi.20961

    doi:10.1002/asi.20961 2009

  26. [26]

    A survey of modern authorship attribution methods,

    E. Stamatatos, “A survey of modern authorship attribution methods,” J. Am. Soc. Inf. Sci. Technol., vol. 60, no. 3, pp. 538–556, Mar. 2009, doi: 10.1002/asi.21001

    doi:10.1002/asi.21001 2009

  27. [27]

    Language models are few-shot learners,

    T. Brown et al., “Language models are few-shot learners,” Adv. Neural Inf. Process. Syst., vol. 33, pp. 1877–1901, 2020

    1901

  28. [28]

    A Survey of Fake News: Fundamental Theories, Detection Methods, and Opportunities,

    X. Zhou and R. Zafarani, “A Survey of Fake News: Fundamental Theories, Detection Methods, and Opportunities,” ACM Comput. Surv., vol. 53, no. 5, pp. 1–40, Sep. 2021, doi: 10.1145/3395046

    doi:10.1145/3395046 2021