Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Evaluation of LLM Vulnerabilities to Being Misused for Personalized Disinformation Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Audience-tailored prompts weaken LLM safety filters and evade detection

desk verdict Useful empirical study of personalized disinformation generation, but the jailbreak claim outruns the data—small untested differences and a known confound. read the letter →

arxiv 2412.13666 v2 pith:OZXCCX6V submitted 2024-12-18 cs.CL cs.AIcs.CY

Aneta Zugecova , Dominik Macko , Ivan Srba , Robert Moro , Jakub Kopal , Katarina Marcincinova , Matus Mesarcik This is my paper · ORCID
classification cs.CLcs.AIcs.CY
keywords personalizeddisinformationjailbreaksafetyfiltersLLMmeta-evaluationmachine-generatedtextdetectionpromptengineeringtarget-grouppersonalizationPerDisNews
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

This paper asks whether current large language models can be steered to produce disinformation tailored to a specific audience, and what that does to the models' built-in safety behavior. Using six LLMs, six disinformation narratives, and seven target groups, the authors generate 2,268 news-style articles under three prompting conditions: no personalization, a simple target-group name, and a detailed audience description. They report that the models mostly produce texts judged well-personalized to the target group, and that adding personalization progressively lowers the rate at which safety filters fire, from 5.2% to 4.5% to 3.5%. They also find that personalized texts are slightly harder for machine-text detectors to spot, with the average true-positive rate falling from 0.9087 to 0.8774. The authors' central claim is that asking an LLM to adapt disinformation to a specific audience functions as a prompt-level jailbreak.

What carries the argument

The key machinery is the PerDisNews dataset construction and its paired evaluation pipeline: six generators, six fact-checker-sourced disinformation narratives, seven target groups defined by European stereotypes, and three personalization-prompt conditions. Personalization quality is scored by a meta-evaluation panel of three different LLMs (GPT-4o, Gemma-2-27b-IT, Llama-3.1-70B-Instruct) whose average is validated against five human annotators on a balanced 109-text subset. Safety-filter activation is detected by a Gemma-based meta-evaluator cross-checked against heuristic phrase matching (97% agreement, Cohen's κ = 0.89). Detectability is measured with three pre-trained detectors (a fine-tuned Gemma-2-9b-IT, Detection-Longformer, and Binoculars) using thresholds calibrated on the original dataset from prior work.

What would settle it

Generate the same narratives under two matched prompt conditions — detailed audience descriptions versus equally long non-personalized filler — and compare safety-filter activation; if activation does not drop in the audience condition, the jailbreak is a prompt-length artifact. Separately, have members of the actual target groups rate the texts; if their ratings do not correlate with the stereotype-based rubric, the personalization-quality claims lose their ground.

Watch

Extended reading notes

Core claim

The paper's central discovery is that personalization itself is a jailbreak vector for disinformation generation. In the authors' experiments, increasing the specificity of the audience description in the prompt — from none, to a target-group name, to a detailed stereotype-based description — steadily reduces how often the tested LLMs refuse to produce the requested disinformation article. The effect holds for every generator individually, even for the model that refused most often overall. The same gradient lowers the detectability of the output: the average true-positive rate across three detection methods drops from 0.9087 (no personalization) to 0.8924 (simple) to 0.8774 (detailed), a statistically significant but small decline. The authors also establish that a three-LLM meta-evaluation of personalization quality correlates strongly with human judgments (Spearman ρ = 0.76 on a 109-text balanced subset), which is the evidence they rely on for the quality ratings that support the jailbreak conclusion.

Load-bearing premise

The 2,268-text quality ratings rest on a 109-text human-validated subset, and both the detailed prompts and the scoring rubrics rely on the same Eurocentric stereotypes about the seven target groups; if those stereotypes misrepresent real audiences, the well-personalized rates and the jailbreak gradient would be inflated.

Editorial extensions

If this is right

  • Safety-filter activation rates of only a few percent mean that most of the tested models will produce requested disinformation articles, and asking for audience tailoring makes refusals even rarer.
  • The jailbreak gradient is monotone in the level of personalization across all six generators, so the effect is not an artifact of one model.
  • Personalized disinformation remains detectable by current detectors in most cases, so detection tools degrade but do not fail.
  • LLM-as-judge evaluation of personalization can scale to thousands of texts while tracking human judgment, enabling larger safety audits.
  • Differences across target groups (conservatives easiest, students and urban residents hardest) suggest that personalization quality varies with how stereotyped the group description is.

Reading between the lines

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

  • The detailed prompts are longer than the no-personalization prompts, so the paper's own limitation note leaves open whether the safety-filter drop is caused by personalization content or simply by prompt length; a control with long but non-personalized instructions would settle this.
  • Because the same target-group descriptions are used both to prompt generation and to score personalization, the quality numbers may partly measure how well models echo the supplied stereotype, not how well the text would persuade real members of those groups; evaluation by actual group representatives is the natural next test.
  • If personalized outputs evade detectors slightly because personalized phrasing is underrepresented in detector training data, then detection benchmarks should start including personalized machine text to avoid a widening gap.
  • The jailbreak effect suggests a concrete mitigation: safety filters should treat audience-adaptation requests as a risk marker and re-check the final text against the disinformation narrative, not just the original prompt.
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

4 major / 5 minor

Summary. The paper constructs PerDisNews, a dataset of 2,268 LLM-generated English disinformation articles, varying six generators, six narratives, seven target groups, and three personalization-prompt conditions (No, Simple, Detailed). It evaluates personalization quality using a three-LLM meta-evaluation validated against five author annotators on 109 texts, measures safety-filter refusal rates, and tests three machine-text detectors on the generated corpus. The main claims are that current LLMs mostly generate well-personalized disinformation, that adding personalization to the prompt reduces safety-filter activation and therefore 'effectively serves as a jailbreak,' that LLM meta-evaluation correlates strongly with human judgment, and that personalization slightly decreases detectability of the generated texts.

Significance. If the claims hold, the paper makes a useful contribution: a relatively large, multi-model dataset for studying personalized disinformation, a scalable LLM-based evaluation protocol with human validation, and a concrete comparison of detection methods on personalized content. The explicit attention to model diversity, the inclusion of open-weight models, and the released analysis code are strengths, as is the attempt to validate the meta-evaluator rather than assuming its reliability. The detectability result is informative even though the effect is small. However, the headline jailbreak claim rests on small percentage differences that are not statistically tested and are acknowledged in the Limitations section to be possibly confounded by prompt length. The meta-evaluation validation is limited to a small, author-annotated subset, and the target-group descriptions used in the detailed-prompt condition double as the scoring rubric, so the personalization-quality numbers partly reflect prompt adherence. These issues do not invalidate the dataset or the meta-evaluation approach, but they require substantial additional analysis before the central claims can be accepted as stated.

major comments (4)
  1. [Section 5.1, Figure 3] The jailbreak claim is not supported by the evidence as reported. The paper reports safety-filter activation of 5.2%, 4.5%, and 3.5% for No, Simple, and Detailed conditions, corresponding to roughly 39, 34, and 26 refusals out of 756 texts per condition, but no significance test, confidence interval, or effect size is provided. The difference between the extremes is small; an approximate two-proportion z-test for 39/756 versus 26/756 does not reach significance at alpha = 0.05. Because the conditions also differ in prompt length and instruction content, and the Limitations section explicitly states that 'other confounding factors (e.g., length of the prompt) might have influenced the results,' the abstract's statement that personalization 'effectively serves as a jailbreak' overstates the finding. The authors should add inferential statistics, control for prompt length, report per-model effects, and revise the abstract and conclusions to match the strength of the evidence.
  2. [Section 5.1 versus Table 3] There is an unresolved inconsistency in the safety-filter counts. The text says Gemma activated safety filters in '152 out of 378' cases, which is 40%, while Table 3 reports a 66% meta-evaluation safety-filter rate for Gemma, corresponding to roughly 249 texts out of 378; Figure 3's aggregate counts also need to be reconciled with these numbers. Without a consistent counting pipeline, readers cannot verify the central descriptive result or reproduce the percentages cited in the abstract.
  3. [Section 3, Appendix C.2, Table 6] The validation of the personalization-quality meta-evaluation is limited in ways that directly affect RQ1. The 109-text human subset is annotated by five of the authors, is balanced according to GPT-4o scores, and excludes safety-filtered texts. More importantly, the detailed-prompt condition supplies the target-group descriptions (Table 6) that are then used as the scoring rubric by both the LLM and human evaluators, so 'well-personalized' scores partly measure adherence to those descriptions rather than effectiveness for actual audiences. The authors should report per-score reliability, show that the meta-evaluation transfers to texts outside the balanced subset, and discuss or test the dependence of the personalization-quality results on the particular target-group descriptions.
  4. [Section 5.3, Tables 4-5] The detectability comparison is confounded by the composition of the prompt conditions. The thresholds are calibrated on the Vykopal et al. dataset and then applied to PerDisNews, and the No, Simple, and Detailed conditions differ in the number of safety-filtered and noisy texts (as established in Section 5.1). If refusal messages are detected at a different rate than full articles, the observed TPR decrease may reflect composition rather than personalization. The authors should report TPR after excluding safety-filtered and noisy texts, and should present the per-generator and per-narrative results for the main comparison rather than only in the appendix.
minor comments (5)
  1. [Section 4.1 and Figure 13] There are typographical errors: 'Flacon' should be 'Falcon' in Section 4.1, and 'Disgreement' should be 'Disagreement' in the Figure 13 label.
  2. [Section 5.2] The statement that human annotators assigned the same score in 33% of cases and majority agreement was 90% would be easier to interpret with confidence intervals and a breakdown by score category; currently the reader cannot assess the reliability of the validation beyond the aggregate percentages.
  3. [Appendix D] The validation correlations for the single-model meta-evaluation on existing datasets (rho = 0.54, 0.69, 0.53) are reported without confidence intervals or significance levels; these should be added to support the claim that the meta-evaluation is usable.
  4. [Figure 3] The percentages cited in the text (5.2%, 4.5%, 3.5%) are not directly readable from the stacked count bars; a small table or axis labels with percentages would improve transparency.
  5. [Section 3] Since the exact prompts are not disclosed for ethical reasons, the generic prompt description should be accompanied by a more explicit statement of which prompt components are withheld and how this affects replication, beyond the current one-sentence note.

Circularity Check

1 steps flagged · score 4.0 of 10

The RQ1 personalization-quality result is partly self-definitional: the Detailed prompt and the evaluator rubric share the same target-group descriptions, while the jailbreak and detectability findings remain independent.

  1. self definitional [Appendix C (Design of Human Study), see also Section 5.1 and Table 6]
    "Annotators were provided with example text snippets for each point on the scale and a description of target group attributes (the same as used in the detailed prompt and summarized in Table 6) to steer their understanding of the target groups."

    The Detailed personalization prompt supplies the LLM with the exact target-group attributes from Table 6. The same attributes are then given to human annotators as the standard for judging how well a text is personalized, and the LLM meta-evaluation is validated against this human rubric before being applied to the full dataset. Therefore, the RQ1 finding that "Detailed specification of target group increases personalization quality" largely measures whether the generated text echoes the description that was already inserted into the prompt; the rubric defines that echo as high-quality personalization. The increase is inflated by construction because the generation input and the evaluation criterion coincide.

full rationale

The most load-bearing circularity is in the personalization-quality evaluation. The paper's Detailed prompt condition and its personalization-quality rubric are built from the same Table 6 target-group descriptions, so a text that reproduces those attributes is likely to be scored as well-personalized. This makes the RQ1 quality comparison partly self-definitional and justifies a moderate score. The jailbreak claim about reduced safety-filter activation is not circular in the same way, because refusal rates are measured independently of the rubric; however, it is confounded by prompt length and the paper's own Limitations section acknowledges that the relationship is not necessarily causal. That is a validity concern rather than a circularity. The meta-evaluation validation is also not circular: balancing the human subset on GPT-4o scores is a sampling design choice, and correlation with human judgment remains evidence rather than a reduction. The self-citations to Vykopal et al. and Macko et al. are methodological reuse and external benchmark selection, not load-bearing circular arguments. Overall, one partial self-definitional step affects the central personalization-quality claim, while the jailbreak and detectability findings retain independent content, so a score of 4 is appropriate.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper introduces no fitted numerical constants and no new theoretical entities. The central claims depend on methodological choices: representativeness of six narratives and seven target groups, the validity of LLM meta-evaluation beyond a 109-text author-annotated subset, and the transfer of detection thresholds from Vykopal et al. data to PerDisNews. These are domain assumptions, listed above.

assumptions (4)
  • domain assumption The six disinformation narratives selected from the 20 in Vykopal et al. (2024) are representative of health and politics disinformation for studying personalization.
    Used to generalize RQ1 and RQ3 findings; selection is by authors' judgment (Section 3).
  • domain assumption The LLM-generated target-group descriptions adequately capture the real target audiences.
    Descriptions were generated by LLMs and filtered manually; they are stereotypes, and they are used as both prompt content and evaluation rubric (Appendix C, Table 6).
  • domain assumption LLM meta-evaluation scores on the full dataset remain valid outside the 109-text human-validated subset.
    Validation correlation rho 0.76 is on a small, author-annotated subset; the paper applies the meta-evaluation to all 2,268 texts (Section 5.2).
  • domain assumption Detection thresholds calibrated on the Vykopal et al. (2024) dataset transfer to PerDisNews without retuning.
    TPR comparisons in Table 5 use thresholds from Table 4; domain shift may affect TPR estimates (Section 5.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluation of LLM Vulnerabilities to Being Misused for Personalized Disinformation Generation." pith.science (2026). https://pith.science/paper/OZXCCX6V

@misc{pith2026241213666,
  author       = {Pith},
  title        = {Pith review of: Evaluation of LLM Vulnerabilities to Being Misused for Personalized Disinformation Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OZXCCX6V}},
  note         = {Machine review of arXiv:2412.13666}
}
read the original abstract

The capabilities of recent large language models (LLMs) to generate high-quality content indistinguishable by humans from human-written texts raises many concerns regarding their misuse. Previous research has shown that LLMs can be effectively misused for generating disinformation news articles following predefined narratives. Their capabilities to generate personalized (in various aspects) content have also been evaluated and mostly found usable. However, a combination of personalization and disinformation abilities of LLMs has not been comprehensively studied yet. Such a dangerous combination should trigger integrated safety filters of the LLMs, if there are some. This study fills this gap by evaluating vulnerabilities of recent open and closed LLMs, and their willingness to generate personalized disinformation news articles in English. We further explore whether the LLMs can reliably meta-evaluate the personalization quality and whether the personalization affects the generated-texts detectability. Our results demonstrate the need for stronger safety-filters and disclaimers, as those are not properly functioning in most of the evaluated LLMs. Additionally, our study revealed that the personalization actually reduces the safety-filter activations; thus effectively functioning as a jailbreak. Such behavior must be urgently addressed by LLM developers and service providers.

Figures

Figures reproduced from arXiv: 2412.13666 by the authors.

Figure 1
Figure 1. Meta-evaluation based personalization-quality [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Meta-evaluation of LLM-generated texts stance towards the disinformation narratives. All LLMs except for Gemma generate mostly texts agreeing with the disinformation narratives. stances. Separate answers of agreement and dis￾agreement as well as aggregation based on individ￾ual narratives are provided in Appendix F. The results show that all the generators except for Gemma are mostly agreeing with the dis￾informatio… view at source ↗
Figure 3
Figure 3. Meta-evaluation scores distribution over the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Meta-evaluation scores distribution over the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Distribution of annotation scores assigned by [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Distribution of meta-evaluation scores as [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Receiver operating characteristic curves of the [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Meta-evaluation of LLM-generated texts score [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Meta-evaluation of LLM-generated texts [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Meta-evaluation of LLM-generated texts disagreement with the disinformation narratives. 0 50 100 150 200 250 300 350 Text Count H1 H2 H3 P1 P2 P3 Narrative Stance towards disinformation narrative Agree Both Disagree [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Meta-evaluation of LLM-generated texts [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Meta-evaluation of LLM-generated texts agreement with individual disinformation narratives. Identification of narratives is based on [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: Meta-evaluation of LLM-generated texts disagreement with individual disinformation narratives. Identification of narratives is based on [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Security Concerns for Large Language Models: A Survey

    cs.CR 2025-05 conditional novelty 5.0 of 10

    A survey that classifies LLM security threats and argues that intrinsic agentic risks, such as scheming, are underappreciated and poorly defended.

Reference graph

Works this paper leans on

46 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [1]

    Rubén Arcos, Manuel Gértrudix, Cristina Arribas, and Monica Cardarilli. 2022. https://doi.org/10.12688/openreseurope.14088.1 Responses to digital disinformation as part of hybrid threats: A systematic review on the effects of disinformation and the effectiveness of fact-checking/debunking . Open Research Europe, 2:8

    work page doi:10.12688/openreseurope.14088.1 2022

  2. [2]

    Hritik Bansal, John Dang, and Aditya Grover. 2023. Peering through preferences: Unraveling feedback acquisition for aligning large language models. arXiv preprint arXiv:2308.15812

    arXiv 2023

  3. [3]

    Dipto Barman, Ziyi Guo, and Owen Conlan. 2024. The dark side of language models: Exploring the potential of llms in multimedia disinformation generation and dissemination. Machine Learning with Applications, page 100545

    work page 2024

  4. [4]

    Judit Bayer, Natalija Bitiukova, Petra Bárd, Alberto Szakácsudit, AlbertoAlemanno, and Erik Uszkiewicz. 2019. https://doi.org/10.2139/ssrn.3409279 Disinformation and propaganda – impact on the functioning of the rule of law in the eu and its member states . SSRN Electronic Journal

    work page doi:10.2139/ssrn.3409279 2019

  5. [5]

    Bennett and Steven Livingston

    Lance W. Bennett and Steven Livingston. 2018. https://doi.org/10.1177/0267323118760317 The disinformation order: disruptive communication and the decline of democratic institutions . European Journal of Communication, 33(2):122--139

    work page doi:10.1177/0267323118760317 2018

  6. [6]

    Jan Blom. 2000. https://doi.org/10.1145/633292.633483 Personalization: a taxonomy . In CHI '00 Extended Abstracts on Human Factors in Computing Systems, CHI EA '00, page 313–314, New York, NY, USA. Association for Computing Machinery

    work page arXiv 2000

  7. [7]

    Ali Borji. 2023. https://arxiv.org/abs/2302.03494 A categorical archive of chatgpt failures . Preprint, arXiv:2302.03494

    arXiv 2023

  8. [8]

    Ben Buchanan, Andrew Lohn, Micah Musser, and Katerina Sedova. 2021. Truth, lies, and automation. Center for Security and Emerging technology, 1(1):2

    work page 2021

Show all 46 references
  1. [9]

    Pengshan Cai, Kaiqiang Song, Sangwoo Cho, Hongwei Wang, Xiaoyang Wang, Hong Yu, Fei Liu, and Dong Yu. 2023. Generating user-engaging news headlines. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3265--3280

    2023

  2. [10]

    Evan N Crothers, Nathalie Japkowicz, and Herna L Viktor. 2023. Machine-generated text: A comprehensive survey of threat models and detection methods. IEEE Access, 11:70977--71002

    2023

  3. [11]

    Saadia Gabriel, Liang Lyu, James Siderius, Marzyeh Ghassemi, Jacob Andreas, and Asu Ozdaglar. 2024. Generative ai in the era of'alternative facts'

    2024

  4. [12]

    Goldstein, Girish Sastry, Micah Musser, Renee DiResta, Matthew Gentzel, and Katerina Sedova

    Josh A. Goldstein, Girish Sastry, Micah Musser, Renee DiResta, Matthew Gentzel, and Katerina Sedova. 2023. https://arxiv.org/abs/2301.04246 Generative language models and automated influence operations: Emerging threats and potential mitigations . Preprint, arXiv:2301.04246

    2023 arXiv

  5. [13]

    Greene, Constance de Saint Laurent, Gillian Murphy, Toby Prike, Karen Hegarty, and Ullrich K

    Ciara M. Greene, Constance de Saint Laurent, Gillian Murphy, Toby Prike, Karen Hegarty, and Ullrich K. H. Ecker. 2023. https://doi.org/10.1027/1016-9040/a000491 Best practices for ethical conduct of misinformation research . European Psychologist, 28(3):139--150

    2023 doi

  6. [14]

    Ziyi Guo. 2024. Online disinformation and generative language models: Motivations, challenges, and mitigations. In Companion Proceedings of the ACM on Web Conference 2024, pages 1174--1177

    2024

  7. [15]

    Kobi Hackenburg and Helen Margetts. 2024. Evaluating the persuasive influence of political microtargeting with large language models. Proceedings of the National Academy of Sciences, 121(24):e2403116121

    2024

  8. [16]

    Rishav Hada, Varun Gumma, Mohamed Ahmed, Kalika Bali, and Sunayana Sitaram. 2024. https://doi.org/10.18653/v1/2024.findings-naacl.148 METAL : Towards multilingual meta-evaluation . In Findings of the Association for Computational Linguistics: NAACL 2024, pages 2280--2298, Mexi...

    2024 doi

  9. [17]

    Abhimanyu Hans, Avi Schwarzschild, Valeriia Cherepanova, Hamid Kazemi, Aniruddha Saha, Micah Goldblum, Jonas Geiping, and Tom Goldstein. 2024. https://arxiv.org/abs/2401.12070 Spotting LLMs with binoculars: Zero-shot detection of machine-generated text . Preprint, arXiv:2401.12070

    2024 arXiv

  10. [18]

    Bakir, and Kalina Bontcheva

    Freddy Heppell, Mehmet E. Bakir, and Kalina Bontcheva. 2024. https://arxiv.org/abs/2402.08467 Lying blindly: Bypassing chatgpt's safeguards to generate hard-to-detect disinformation claims at scale . Preprint, arXiv:2402.08467

    2024 arXiv

  11. [19]

    Andrew Heywood. 2008. Politologie, 3 edition. Vydavatelstv \' a nakladatelstv \' Ale s C en e k

    2008

  12. [20]

    Andreas Jungherr, Gonzalo Rivero, and Daniel Gayo-Avello. 2020. Retooling Politics: How Digital Media Are Shaping Democracy. Cambridge University Press

    2020

  13. [21]

    Yafu Li, Qintong Li, Leyang Cui, Wei Bi, Zhilin Wang, Longyue Wang, Linyi Yang, Shuming Shi, and Yue Zhang. 2024. https://doi.org/10.18653/v1/2024.acl-long.3 MAGE : Machine-generated text detection in the wild . In Proceedings of the 62nd Annual Meeting of the Association for ...

    2024 doi

  14. [22]

    Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, et al. 2022. Holistic evaluation of language models. arXiv preprint arXiv:2211.09110

    2022 arXiv

  15. [23]

    Dominik Macko, Jakub Kopal, Robert Moro, and Ivan Srba. 2024 a . https://arxiv.org/abs/2406.12549 Multisocial: Multilingual benchmark of machine-generated text detection of social-media texts . Preprint, arXiv:2406.12549

    2024 arXiv

  16. [24]

    Dominik Macko, Robert Moro, Adaku Uchendu, Jason Lucas, Michiharu Yamashita, Mat \'u s Pikuliak, Ivan Srba, Thai Le, Dongwon Lee, Jakub Simko, and Maria Bielikova. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.616 MULTIT u DE : Large-scale multilingual machine-generated te...

    2023 doi

  17. [25]

    Dominik Macko, Robert Moro, Adaku Uchendu, Ivan Srba, Jason S Lucas, Michiharu Yamashita, Nafis Irtiza Tripto, Dongwon Lee, Jakub Simko, and Maria Bielikova. 2024 b . https://aclanthology.org/2024.findings-emnlp.369 Authorship obfuscation in multilingual machine-generated text...

    2024

  18. [26]

    SC Matz, JD Teeny, Sumer S Vaid, H Peters, GM Harari, and M Cerf. 2024. The potential of generative ai for personalized persuasion at scale. Scientific Reports, 14(1):4692

    2024

  19. [27]

    Marlene Mauk and Max Grömping. 2023. https://doi.org/10.1177/00104140231193008 Online disinformation predicts inaccurate beliefs about election fairness among both winners and losers . Comparative Political Studies, 57(6):965--998

    2023 doi

  20. [28]

    Elyas Meguellati, Lei Han, Abraham Bernstein, Shazia Sadiq, and Gianluca Demartini. 2024. How good are llms in generating personalized advertisements? In Companion Proceedings of the ACM on Web Conference 2024, pages 826--829

    2024

  21. [29]

    Brent Mittelstadt. 2017. https://doi.org/10.1007/s13347-017-0253-7 From individual to group privacy in big data analytics . Philosophy and Technology, 30(4):475--494

    2017 doi

  22. [30]

    OECD . 2022. https://doi.org/10.1787/76972a4a-en Building Trust and Reinforcing Democracy: Preparing the Ground for Government Action . OECD Public Governance Reviews. OECD Publishing, Paris

    2022 doi

  23. [31]

    Arjun Panickssery, Samuel R Bowman, and Shi Feng. 2024. Llm evaluators recognize and favor their own generations. arXiv preprint arXiv:2404.13076

    2024 arXiv

  24. [32]

    Marta Peréz-Escolar, Darren Lilleker, and Alejandro Tapie-Frade. 2023. https://doi.org/10.17645/mac.v11i2.6453 A systematic literature review of the phenomenon of disinformation and misinformation . Media and Communication, 11(2)

    2023 doi

  25. [33]

    Almog Simchon, Matthew Edwards, and Stephan Lewandowsky. 2024. The persuasive effects of political microtargeting in the age of generative artificial intelligence. PNAS nexus, 3(2):pgae035

    2024

  26. [34]

    Felix M Simon, Sacha Altay, and Hugo Mercier. 2023. Misinformation reloaded? fears about the impact of generative ai on misinformation are overblown. Harvard Kennedy School Misinformation Review, 4(5)

    2023

  27. [35]

    Michal Spiegel and Dominik Macko. 2024 a . https://doi.org/10.18653/v1/2024.acl-demos.17 IMGTB : A framework for machine-generated text detection benchmarking . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonst...

    2024 doi

  28. [36]

    Michal Spiegel and Dominik Macko. 2024 b . https://doi.org/10.18653/v1/2024.semeval-1.84 KI n IT at S em E val-2024 task 8: Fine-tuned LLM s for multilingual machine-generated text detection . In Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-20...

    2024 doi

  29. [37]

    Gemma Team. 2024. https://doi.org/10.34740/KAGGLE/M/3301 Gemma

    2024 doi

  30. [38]

    Yuliia Turchenko, Kira Horiacheva, Oleksandr Dzhus, and Oleh Kolisnyk. 2021. https://doi.org/10.2478/kbo-2021-0078 Disinformation as a threat to the quality of contemporary information . International Conference KNOWLEDGE-BASED ORGANIZATION, 27(2):225--228

    2021 doi

  31. [39]

    Ivan Vykopal, Mat \'u s Pikuliak, Ivan Srba, Robert Moro, Dominik Macko, and Maria Bielikova. 2024. https://doi.org/10.18653/v1/2024.acl-long.793 Disinformation capabilities of large language models . In Proceedings of the 62nd Annual Meeting of the Association for Computation...

    2024 doi

  32. [40]

    Yaqing Wang, Jiepu Jiang, Mingyang Zhang, Cheng Li, Yi Liang, Qiaozhu Mei, and Michael Bendersky. 2023. https://arxiv.org/abs/2310.11593 Automated evaluation of personalized text generation using large language models . Preprint, arXiv:2310.11593

    2023 arXiv

  33. [41]

    Yuxia Wang, Artem Shelmanov, Jonibek Mansurov, Akim Tsvigun, Vladislav Mikhailov, Rui Xing, Zhuohan Xie, Jiahui Geng, Giovanni Puccetti, Ekaterina Artemova, Jinyan Su, Minh Ngoc Ta, Mervat Abassy, Kareem Elozeiri, Saad El Dine Ahmed, Maiya Goloburda, Tarek Mahmoud, Raj Vardhan...

    2025

  34. [42]

    Williams, Liam Burke-Moore, Ryan Sze-Yin Chan, Florence E

    Angus R. Williams, Liam Burke-Moore, Ryan Sze-Yin Chan, Florence E. Enock, Federico Nanni, Tvesha Sippy, Yi-Ling Chung, Evelina Gabasova, Kobi Hackenburg, and Jonathan Bright. 2024. https://arxiv.org/abs/2408.06731 Large language models can consistently generate high-quality c...

    2024 arXiv

  35. [43]

    Wanzheng Zhu and Suma Bhat. 2020. https://doi.org/10.18653/v1/2020.findings-emnlp.9 GRUEN for evaluating linguistic quality of generated text . In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 94--108, Online. Association for Computational Linguistics

    2020 doi

  36. [44]

    Terry Yue Zhuo, Yujin Huang, Chunyang Chen, and Zhenchang Xing. 2023. https://arxiv.org/abs/2301.12867 Red teaming chatgpt via jailbreaking: Bias, robustness, reliability and toxicity . Preprint, arXiv:2301.12867

    2023 arXiv

  37. [45]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  38. [46]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.