Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Tailored Truths: Optimizing LLM Persuasion with Personalization and Fabricated Statistics

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read The paper claims that a GPT-4o-mini multi-agent debate strategy using personalized, fabricated statistics persuades 51% of participants to move their opinion, compared with 32% for a static human-written argument.

desk verdict The 51%-vs-32% headline is not statistically reliable, but the paper's transparency and the cheap-fabrication demonstration make it worth a proper review. read the letter →

arxiv 2501.17273 v1 pith:7WDWKPME submitted 2025-01-28 cs.CL

classification cs.CL
keywords LLMpersuasionmicrotargetingfabricatedstatisticsmulti-agentdebateopinionchangedisinformationpersonalizedargumentsGPT-4o-mini
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 tries to establish how persuasive current, cheap large language models can be when they argue interactively with a person, and whether simple scaffolding can push them past human-written text. Across 198 debates with 33 participants, it finds that a three-agent debate setup that combines personal details with fabricated statistics changed the intended opinion 51% of the time, versus 32% for static human arguments and about 46% for static LLM arguments. A plain personalized prompt, by contrast, performed worse than the model's default debate style. The authors read this as evidence that interactive, tailored, statistic-laced LLM arguments are a realistic and very cheap tool for large-scale disinformation, and they include cost estimates comparing such debates to human-run influence operations. The study is small and noisy, so the numerical gap should be read cautiously.

What carries the argument

The key machinery is the mixed multi-agent debate scaffold. Three GPT-4o-mini agents share a private chat before each reply: a personalization agent reads the user's demographics and Big Five personality traits, a statistics agent drafts realistic-looking made-up data with plausible sources, and an executive agent composes the final 200-word response. The paper also uses a fixed three-phase debate format (introduction, rebuttal, conclusion) in which the participant argues first for their initial side, and measures persuasion as Likert-scale shift pre and post. The machinery's work is to let the model focus on one or two traits and one or two invented statistics instead of diluting its argument, which the paper argues explains why the mixed type beat both the personalized-only and stats-only types.

What would settle it

Run a preregistered replication with n=300 per condition using the same topics and protocol; if the 95% confidence interval for the mixed minus human-static difference in the probability of shifting opinion contains zero, the central claim fails. A cheaper check is a paired permutation test on the raw participant-level data from the published dataset.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the persuasion method matters more than the source: GPT-4o-mini's static arguments were comparable to human-written ones, but the same model in a three-phase debate was 'meaningfully more persuasive' when an executive agent synthesized the advice of a personalization agent and a statistics agent. This mixed strategy yielded an estimated marginal mean Likert shift of 1.146 and a 51% probability of moving the participant in the intended direction, while human static arguments yielded 0.833 and 32.1%. The paper interprets the result as showing that a specific combination—personalization plus fabricated statistics, coordinated through a multi-agent scratchpad—is what unlocks the model's persuasive advantage, rather than either ingredient alone.

Load-bearing premise

The central claim rests on the assumption that the 19-point gap between the mixed and human-static conditions is a real signal rather than sampling noise, in a 33-person study whose confidence intervals overlap and whose p-values range from 0.276 to 0.976.

Editorial extensions

If this is right

  • If the 51% result holds, a single debate costing about $0.00449 can move more than half of engaged participants on non-partisan topics, making LLM-driven influence operations dramatically cheaper than human-run ones.
  • The result implies that access to basic demographics and a short personality test is enough to materially boost a scaffolded LLM's persuasion, since the mixed strategy used exactly that data.
  • Because static LLM arguments were not clearly better than human ones, the paper implies that the danger is not better writing but interactive tailoring at scale.
  • The finding also implies that outright fabrication can be part of a winning strategy; the LLM was prompted to invent statistics, and the mixed condition still outperformed the honest default.

Reading between the lines

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

  • My read: the forced debate format may contribute to the measured shift, since participants must argue for their initial side before reading the rebuttal; a free-form or social-media-style replication could show a smaller gap.
  • A natural extension the paper does not run: replace fabricated statistics with true statistics while keeping the multi-agent scaffold; if the 51% effect persists, the persuasion comes from presentation and targeting rather than deception.
  • The regression results suggest that gender and whether the LLM argues for or against the statement mattered more than any personality trait; future work could test whether personalization is actually doing the work or whether the executive agent's scratchpad alone explains the gain.
  • If the cost arithmetic generalizes, defensive systems may need to match this interactivity: static warning labels or detection will likely underperform interactive counter-messaging.
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 / 6 minor

Summary. This paper reports a human-participant study (n=33, m=198 interactions) comparing the persuasiveness of GPT-4o-mini in structured debates under four strategies—simple, statistics-focused, personalized, and a mixed multi-agent approach combining personalization and fabricated statistics—against static human- and LLM-written arguments. Persuasiveness is measured by change on a 7-point Likert scale and by the probability of an opinion shift in the intended direction. The authors report that the mixed debate strategy had a 51% probability of changing participants' positions versus 32% for static human arguments, and conclude that LLM-based mixed strategies are 'meaningfully more persuasive' and pose a scalable disinformation risk. The paper provides open-source code, detailed prompts, cost estimates, and an explicit limitations section.

Significance. If the headline comparison were statistically reliable, this would be a useful contribution to the growing literature on LLM persuasion and would strengthen policy concerns about scalable disinformation. The paper's strengths include its transparent reporting of prompts and platform design, its decision to measure both magnitude and direction of opinion change, its debriefing and consent procedures for the use of fabricated statistics, and a rare cost-per-debate analysis that makes the threat scenario concrete. However, the main comparative claim is currently not supported by the inferential statistics reported in the manuscript, so the significance is prospective rather than established.

major comments (4)
  1. [Section 4, Table 2, Appendix G.1] The central claim that the mixed condition is 'meaningfully more persuasive' than static human arguments rests on a 19-point difference in P(+change) (51.0±17.0 vs 32.1±16.3) whose 95% confidence intervals overlap substantially; the paper's own report of p-values between 0.276 and 0.976 and R²=0.065 for the main regression indicates that the observed difference cannot be distinguished from sampling noise. Because the abstract, introduction, and conclusion all present the 51% vs 32% contrast without this caveat, the wording overstates what the data establish; the authors should either reanalyze with a pre-specified mixed-effects model and report adjusted pairwise tests, or revise the abstract and conclusions to describe the result as exploratory.
  2. [Section 4 and Appendix E] The statistical model underlying Table 2 is not fully specified: 198 observations are clustered within 33 participants and across 29 topics, yet the text describes a 'standard linear regression' and a binomial logistic regression without stating whether participant and topic random effects or cluster-robust standard errors were used. Without such correction, the reported confidence intervals are likely to be too narrow, and the headline comparison is even less secure than Table 2 suggests.
  3. [Section 5.1 and Section 6] The mixed condition differs from the simple, stats, and personalized conditions along several dimensions simultaneously—multi-agent scaffolding, a scratchpad, combined personalization and fabricated statistics—so the design cannot identify which component drives the improvement. The conclusion that 'employing a novel multi-agent debate approach... proved to be effective' overstates attribution; the authors acknowledge the conflation in Section 5.1 but should not state component-level conclusions without a decomposition experiment or appropriate caveats in the conclusion.
  4. [Abstract and Section 4] The claim that static arguments generated by humans and GPT-4o-mini have 'comparable persuasive power' is not supported by the reported statistics: arg-hum and arg-llm show opposite ordering on the two metrics (Likert Δ 0.833 vs 0.721; P(+change) 32.1% vs 45.9%), and both confidence intervals overlap. Without formal equivalence testing or a clear statement that the data are inconclusive on this comparison, the equivalence language should be moderated.
minor comments (6)
  1. [Figure placement] Figure 5 is referenced in Section 4 before Figures 3 and 4 are introduced; the figures should be renumbered or reordered so that the first reference matches the numerical order.
  2. [Section 5.1] The text refers to 'm=198 debates' although 66 of the 198 interactions are static argument readings rather than debates; 'interactions' should be used consistently.
  3. [Appendix G] Appendix G is titled 'Other Implications of Lower Costs' but Section G.1 is the extended limitations discussion referenced from the main text; the appendix should be retitled or split so that the limitations material is easy to locate.
  4. [Appendix G.1] The statement that 'p-values for debate types in the model range from 0.276 to 0.976' does not identify which model or metric the p-values refer to; this should be clarified.
  5. [Section 4] The sentence about 'the two highest Likert Δ values of five' coming from the arg-hum type is anecdotal and should be removed or replaced with a quantile-based characterization of the distributions.
  6. [General] The manuscript contains contractions in formal prose (e.g., 'It's notable') and inconsistent use of 'debate' versus 'interaction'; a careful copyedit would improve readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical comparisons of human trial outcomes, not derivations that reduce to their own inputs.

full rationale

This paper contains no mathematical derivation whose output is equivalent to its inputs by construction. The central claims, such as the mixed debate type reaching 51% vs 32% P(+change), are estimated marginal means from regressions on participants' own pre- and post-debate Likert-scale ratings (Section 4, Table 2); the outcome variable is defined independently of the treatment prompts. No parameter is fitted to a subset of data and then reported as a prediction of a closely related quantity; the EMMs are adjusted summaries of the full fitted model. The authors cite Durmus et al. for topic selection and for the observation that fabricated statistics can be persuasive, but that is external evidence, not self-citation, and it informs prompt design rather than defining the measured outcome. The only potentially self-referential element is that GPT-4o-mini was used both to generate arguments and to review participant responses for quality (Section 3.2); however, the measured outcomes are participants' own Likert ratings, and the model review only filters incoherent or off-topic responses, so the outcome is not defined by the model's judgment. The paper explicitly reports overlapping confidence intervals, non-significant p-values (0.276-0.976), R^2=0.065, and small n=33 (Sections 5.1 and G.1), which are statistical power and robustness concerns rather than circularity. No circular step can be quoted with a specific reduction, so the appropriate finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The study is empirical and uses no mathematical derivation, so there are no invented entities and only standard statistical modeling assumptions. The main free parameters are the regression coefficients whose estimated marginal means carry the paper's comparisons.

free parameters (1)
  • Linear regression coefficients for interaction types and covariates = E.g., mixed Likert delta EMM 1.146, P(+change) 51%; simple 0.782/42.7%
    Estimated from 198 trials to obtain the persuasion effect sizes that the paper's comparison relies on. These are fitted to the data, not derived from theory.
assumptions (3)
  • domain assumption Self-reported Likert scale shifts measure persuasion
    Section 4 defines Likert delta as the outcome; assumes one-point moves reflect genuine opinion change rather than demand characteristics or fatigue.
  • domain assumption The pooled static human arguments from Durmus et al. and new Prolific writing are a valid baseline
    Section 3.1 combines external dataset with newly collected arguments; assumes comparability despite different elicitation methods (noted in Appendix G).
  • domain assumption Randomized order of six conditions within a 1h45m session controls for order and fatigue effects
    Section 3.2 randomizes interaction order; assumes no differential carryover, though fatigue is acknowledged.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tailored Truths: Optimizing LLM Persuasion with Personalization and Fabricated Statistics." pith.science (2026). https://pith.science/paper/7WDWKPME

@misc{pith2026250117273,
  author       = {Pith},
  title        = {Pith review of: Tailored Truths: Optimizing LLM Persuasion with Personalization and Fabricated Statistics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7WDWKPME}},
  note         = {Machine review of arXiv:2501.17273}
}
abstract

Large Language Models (LLMs) are becoming increasingly persuasive, demonstrating the ability to personalize arguments in conversation with humans by leveraging their personal data. This may have serious impacts on the scale and effectiveness of disinformation campaigns. We studied the persuasiveness of LLMs in a debate setting by having humans $(n=33)$ engage with LLM-generated arguments intended to change the human's opinion. We quantified the LLM's effect by measuring human agreement with the debate's hypothesis pre- and post-debate and analyzing both the magnitude of opinion change, as well as the likelihood of an update in the LLM's direction. We compare persuasiveness across established persuasion strategies, including personalized arguments informed by user demographics and personality, appeal to fabricated statistics, and a mixed strategy utilizing both personalized arguments and fabricated statistics. We found that static arguments generated by humans and GPT-4o-mini have comparable persuasive power. However, the LLM outperformed static human-written arguments when leveraging the mixed strategy in an interactive debate setting. This approach had a $\mathbf{51\%}$ chance of persuading participants to modify their initial position, compared to $\mathbf{32\%}$ for the static human-written arguments. Our results highlight the concerning potential for LLMs to enable inexpensive and persuasive large-scale disinformation campaigns.

Figures

Figures reproduced from arXiv: 2501.17273 by the authors.

Figure 1
Figure 1. Diagram depicting the process used to generate the Mixed approach responses. The messages seen here are excerpts from one interaction recorded during our experiments. selected from the 29 topics and displayed to the user, who was then asked to briefly describe their agreement with the statement. 2. The user reported how strongly they agreed or dis￾agreed with the premise of the topic on a Likert scale from one to se… view at source ↗
Figure 2
Figure 2. Diagram describing the process flow for each interaction [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Likert ∆ adjusted mean (EMM) by argument and debate type magnitude, measured as the proportion of cases where Likert ∆ is greater than zero. To analyze these metrics, we conducted a standard linear re￾gression for Likert ∆ and a binomial logistic regression for P(+change), both comparing across interaction types. Be￾cause interaction type is ordinal, regression coefficients are typically reported relative to a refer… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Violin plot of raw Likert ∆ values by argument and debate type other. This was especially helpful for the personalized agent, allowing it to focus on one or two key traits, a strat￾egy shown earlier to enhance effectiveness. In contrast, the personalized debate type re…

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. Analysing Differences in Persuasive Language in LLM-Generated Text: Uncovering Stereotypical Gender Patterns

    cs.CL 2026-01 conditional novelty 6.0 of 10

    LLMs consistently generate more emotional/communal persuasion for female targets and more direct/agentic persuasion for male targets across models and languages.

Reference graph

Works this paper leans on

52 extracted references · 39 canonical work pages · cited by 1 Pith paper

  1. [1]

    What is israeli firm stoic and how it tried to disrupt lok sabha polls 2024

    Anand, N. What is israeli firm stoic and how it tried to disrupt lok sabha polls 2024. jun 2024. URL https://www.business-standard.com/ elections/lok-sabha-election/openai- report-on-lok-sabha-polls-zero-zeno- what-is-israeli-firm-stoic-and-how-it- tried-to-disrupt-lok-sabha-polls-2024- 124060100518 1.html

  2. [2]

    Challenges in Red Teaming AI Systems, June

    Anthropic. Challenges in Red Teaming AI Systems, June

  3. [3]

    C., Elson, M., and Schneider, I

    Anvari, F., Efendic, E., Olsen, J., Arslan, R. C., Elson, M., and Schneider, I. K. Bias in self-reports: An initial ele- vation phenomenon. Social Psychological and Personal- ity Science, 2023. URL https://doi.org/10.1177/ 19485506221129160

  4. [4]

    The dark side of language models: Exploring the potential of llms in multimedia disinformation generation and dissemination

    Barman, D., Guo, Z., and Conlan, O. The dark side of language models: Exploring the potential of llms in multimedia disinformation generation and dissemination. Machine Learning with Appli- cations, 16:100545, 2024. ISSN 2666-8270. doi: https://doi.org/10.1016/j.mlwa.2024.100545. URL https://www.sciencedirect.com/science/ article/pii/S2666827024000215

  5. [5]

    Bradshaw, S., Bailey, H., and Howard, P. N. In- dustrialized disinformation: 2020 global inventory of organised social media manipulation. Working Paper 2021.1, Project on Computational Propaganda, Oxford Internet Institute, University of Oxford, Oxford, UK,

  6. [6]

    M., Egdal, D

    Breum, S. M., Egdal, D. V ., Mortensen, V . G., Møller, A. G., and Aiello, L. M. The persuasive power of large lan- guage models, 2023. URL https://arxiv.org/abs/ 2312.15523

  7. [7]

    Truth, Lies, and Automation How Language Models Could Change Disinformation

    Buchanan, J., Lohn, J., Musser, J., and Sedova, J. Truth, Lies, and Automation How Language Models Could Change Disinformation. Technical report, Center for Security and Emerging Technology,Georgetown Univ., 2021

  8. [8]

    Debate format: Three rounds of structured debate, 2020

    Burgos, S. Debate format: Three rounds of structured debate, 2020. URL https://www.mcgill.ca/tls/ files/tls/burgos-debate-structure .pdf. Ac- cessed January 8, 2025

Show all 52 references
  1. [9]

    Non-determinism in gpt-4, 2023

    Chann, S. Non-determinism in gpt-4, 2023. URL https://152334h.github.io/blog/non- determinism-in-gpt-4/ . Accessed: 2025-01- 21

  2. [10]

    Committee on national security systems (cnss) glos- sary

    CNSS. Committee on national security systems (cnss) glos- sary. 2022. URL https://www.cnss.gov/CNSS

  3. [11]

    Durably reducing conspiracy beliefs through dialogues with ai

    Costello, H., Pennycook, G., and Rand, D. Durably reducing conspiracy beliefs through dialogues with ai. Science, 2024. URL https://doi.org/adq1814. 10 Tailored Truths

  4. [12]

    A Guide to Min- imum Wage in India, September 2024

    Dezan Shira and Associates. A Guide to Min- imum Wage in India, September 2024. URL https://www.india-briefing.com/news/ guide-minimum-wage-india-19406 .html/

  5. [13]

    Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., Goyal, A., Hartshorn, A., Yang, A., Mitra, A., Sravankumar, A., Korenev, A., Hinsvark, A., Rao, A., Zhang, A., Rodriguez, A., Gregerson, A., Spataru, A., Rozie...

  6. [14]

    Measuring the persuasiveness of language models

    Durmus, E., Lovitt, L., Tamkin, A., Ritchie, S., Clark, J., and Ganguli, D. Measuring the persuasiveness of language models. 2024. URL https://www.anthropic.com/ news/measuring-model-persuasiveness

  7. [15]

    A day in the life of a Shwe Kokko scammer, June

    Frontier. A day in the life of a Shwe Kokko scammer, June

  8. [16]

    A survey on offensive ai within cybersecurity, 2024

    Girhepuje, S., Verma, A., and Raina, G. A survey on offensive ai within cybersecurity, 2024. URL https: //arxiv.org/abs/2410.03566

  9. [17]

    Predicting personality with social media

    Golbeck, J., Robles, C., and Turner, K. Predicting personality with social media. 2011. 11 Tailored Truths

  10. [18]

    A., Sastry, G., Musser, M., DiResta, R., Gentzel, M., and Sedova, K

    Goldstein, J. A., Sastry, G., Musser, M., DiResta, R., Gentzel, M., and Sedova, K. Generative language models and auto- mated influence operations: Emerging threats and poten- tial mitigations, 2023. URL https://arxiv.org/abs/ 2301.04246

  11. [19]

    D., Rentfrow, P

    Gosling, S. D., Rentfrow, P. J., and Jr., W. B. S. A very brief measure of the big-five personality domains. Journal of Research in Personality , 37:504–528, 2003. doi: 10 .1016/S0092-6566(03)00046-1. URL https:// doi.org/10.1016/S0092-6566(03)00046-1

  12. [20]

    Why report estimated marginal means?, aug 2021

    Grace-Martin, K. Why report estimated marginal means?, aug 2021. URL https: //www.theanalysisfactor.com/why-report- estimated-marginal-means-in-spss-glm/

  13. [21]

    and Margetts, H

    Hackenburg, K. and Margetts, H. Evaluating the persua- sive influence of political microtargeting with large lan- guage models. Proceedings of the National Academy of Sciences, 121(24):e2403116121, 2024. doi: 10 .1073/ pnas.2403116121. URL https://doi.org/10.1073/ pnas.2403116121

  14. [22]

    A russian bot farm used ai to lie to americans

    Harding, E. A russian bot farm used ai to lie to americans. what now?, 2024. URL https://www.csis.org. Pub- lished July 16, 2024

  15. [23]

    Bots and misinformation spread on social media: Implications for covid-19

    Himelein-Wachowiak, M., Giorgi, S., Devoto, A., Rahman, M., Ungar, L., Schwartz, H., Epstein, D., Leggio, L., and Curtis, B. Bots and misinformation spread on social media: Implications for covid-19. J Med Internet Res, 23(5):e26933, may 2021. doi: 10 .2196/26933

  16. [24]

    Of 2 minds: How fast and slow thinking shape perception and choice [excerpt]

    Kahneman, D. Of 2 minds: How fast and slow thinking shape perception and choice [excerpt]. Scientific American, 2012. URL https://www.scientificamerican.com/ article/kahneman-excerpt-thinking-fast- and-slow/

  17. [25]

    Private traits and attributes are predictable from digital records of human behavior

    Kosinski, M., Stillwell, D., and Graepel, T. Private traits and attributes are predictable from digital records of human behavior. Proceedings of the National Academy of Sciences, 2013

  18. [26]

    Remodetect: Reward models recognize aligned llm’s generations, 2024

    Lee, H., Tack, J., and Shin, J. Remodetect: Reward models recognize aligned llm’s generations, 2024. URL https: //arxiv.org/abs/2405.17382

  19. [27]

    and Warren, P

    Linvill, D. and Warren, P. Digital yard signs: Analysis of an ai bot political influence campaign on x. Technical report, Media Forensics Hub,Clemson University, 2024. URL https://open.clemson.edu/mfh reports/7

  20. [28]

    F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., and Liang, P

    Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., and Liang, P. Lost in the middle: How language models use long contexts, 2023. URL https: //arxiv.org/abs/2307.03172

  21. [29]

    ConspEmoLLM: Conspiracy Theory Detection Using an Emotion-Based Large Language Model

    Liu, Z., Liu, B., Thompson, P., Yang, K., and Ananiadou, S. ConspEmoLLM: Conspiracy Theory Detection Using an Emotion-Based Large Language Model. IOS Press, October

  22. [30]

    Malhotra, A., Totti, L., au2, W. M. J., Kumaraguru, P., and Almeida, V . Studying user footprints in different online social networks, 2013. URLhttps://arxiv.org/abs/ 1301.6870

  23. [31]

    C., Teeny, J

    Matz, S. C., Teeny, J. D., Vaid, S. S., Peters, H., Harari, G. M., and Cerf, M. The potential of generative ai for personalized persuasion at scale. Scientific Reports, 14:4692, 2024. doi: 10.1038/s41598-024-53755-0. URL https://doi.org/ 10.1038/s41598-024-53755-0 . Received: ...

  24. [32]

    The dark patterns of personalized persuasion in large language models: Exposing persuasive linguistic features for big five personality traits in llms responses, 2024

    Mieleszczenko-Kowszewicz, W., Płudowski, D., Kołodziejczyk, F., ´Swistak, J., Sienkiewicz, J., and Biecek, P. The dark patterns of personalized persuasion in large language models: Exposing persuasive linguistic features for big five personality traits in llms responses, 2024....

  25. [33]

    doi: 10.3233/faia241060

    ISBN 9781643685489. doi: 10.3233/faia241060. URL http://dx.doi.org/10.3233/FAIA241060

  26. [34]

    Reality Check Commentary: Krem- lin’s World-Class Dashboard Maximizes Disinfor- mation, at 26 Cents Per Lie, February 2024

    NewsGuard. Reality Check Commentary: Krem- lin’s World-Class Dashboard Maximizes Disinfor- mation, at 26 Cents Per Lie, February 2024. URL https://www.newsguardrealitycheck.com/p/ reality-check-commentary-kremlins

  27. [35]

    J., Gur-Ari, G., Michalewski, H., Austin, J., Bieber, D., Dohan, D., Lewkowycz, A., Bosma, M., Luan, D., Sutton, C., and Odena, A

    Nye, M., Andreassen, A. J., Gur-Ari, G., Michalewski, H., Austin, J., Bieber, D., Dohan, D., Lewkowycz, A., Bosma, M., Luan, D., Sutton, C., and Odena, A. Show your work: Scratchpads for intermediate computation with lan- guage models, 2021. URL https://arxiv.org/abs/ 2112.00114

  28. [36]

    OpenAI Red Teaming Network, September 2023

    OpenAI. OpenAI Red Teaming Network, September 2023. URL https://openai.com/index/red-teaming- network/

  29. [37]

    M., De Francisci Morales, G., and Bonchi, F

    Monti, C., Aiello, L. M., De Francisci Morales, G., and Bonchi, F. The language of opinion change on social media under the lens of communicative action. Scientific Reports, 12(1), October 2022. ISSN 2045-2322. doi: 10.1038/s41598- 022-21720-4. URL http://dx.doi.org/10.1038/ s...

  30. [38]

    Disrupting deceptive uses of ai by covert influence operations, May 2024

    OpenAI. Disrupting deceptive uses of ai by covert influence operations, May 2024. URL https://openai.com/ index/disrupting-deceptive-uses-of-AI- by-covert-influence-operations/

  31. [39]

    O’Reilly-Shah, V . N. Factors influencing healthcare provider respondent fatigue answering a globally administered in-app survey. PeerJ, 5, 2017. URL https://api.semanticscholar.org/CorpusID: 23555065

  32. [40]

    Park, G. et al. Automatic personality assessment through social media language. Journal of Personality and Social Psychology, 2015

  33. [41]

    Api pricing for openai models, 2024

    OpenAI. Api pricing for openai models, 2024. URL https: //openai.com/api/pricing/. Accessed: 2024-11- 11

  34. [42]

    H., Gallotti, R., and West, R

    Salvi, F., Ribeiro, M. H., Gallotti, R., and West, R. On the conversational persuasiveness of large language mod- els: A randomized controlled trial, 2024. URL https: //arxiv.org/abs/2403.14380. 12 Tailored Truths

  35. [43]

    Quantifying the potential persuasive returns to political microtargeting

    Tappin, B., Wittenberg, C., Hewitt, L., Berinsky, A., and Rand, D. Quantifying the potential persuasive returns to political microtargeting. Proceedings of the National Academy of Sciences , 120(25):e2216261120, 2023. doi: 10.1073/pnas.2216261120. URL https://doi.org/ 10.1073/...

  36. [44]

    Government Accountability Office

    U.S. Government Accountability Office. High-risk series: Urgent action needed to address critical cybersecurity chal- lenges facing the nation. Technical Report GAO-24-107231, jun 2024

  37. [45]

    Paulhus, D. L. and Vazire, S. The self-report method. In Robins, R. W., Fraley, R. C., and Krueger, R. F. (eds.),Hand- book of Research Methods in Personality Psychology , pp. 224–239. The Guilford Press, 2007

  38. [46]

    Can chatgpt defend its belief in truth? evaluating llm reasoning via debate, 2023

    Wang, B., Yue, X., and Sun, H. Can chatgpt defend its belief in truth? evaluating llm reasoning via debate, 2023. URL https://arxiv.org/abs/2305.13160

  39. [47]

    Mixture-of-agents enhances large language model capabilities, 2024

    Wang, J., Wang, J., Athiwaratkun, B., Zhang, C., and Zou, J. Mixture-of-agents enhances large language model capabilities, 2024. URL https://arxiv.org/abs/ 2406.04692

  40. [48]

    workers at Russian troll farms earn 660 USD equivalent per month for writing 100 comments per day on social media

    Woollacott, E. Top ai chatbots spread russian propaganda. Forbes, 2024. URL https://www.forbes.com/ sites/emmawoollacott/2024/06/19/top-ai- chatbots-spread-russian-propaganda/ . 13 Tailored Truths A. Human Participants We used Prolific’s built-in pre-study screening filters to...

  41. [49]

    Who does(n’t) target you? mapping the worldwide usage of online political microtar- geting

    V otta, F., Kruschinski, S., Hove, M., Helberger, N., Dob- ber, T., and de Vreese, C. Who does(n’t) target you? mapping the worldwide usage of online political microtar- geting. Journal of Quantitative Description: Digital Me- dia, 4, 2024. doi: 10 .51685/jqd.2024.010. URL htt...

  42. [2021]

    This work is licensed under a Creative Commons Attribution - Non Commercial - Share Alike 4.0 International License

    URL https://www.oii.ox.ac.uk/research/ projects/computational-propaganda/. This work is licensed under a Creative Commons Attribution - Non Commercial - Share Alike 4.0 International License

  43. [2023]

    URL https://www.frontiermyanmar.net/ en/a-day-in-the-life-of-a-shwe-kokko- scammer/

  44. [2024]

    URL https://www.anthropic.com/news/ challenges-in-red-teaming-ai-systems

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.