Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

This paper claims that LLM-based agents playing Werewolf change their night targets, votes, and trust judgments when gender is revealed, and that first names alone produce similar discriminatory behavior.

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 ↗

In Werewolf games, LLM agents change their kills, votes, and trust scores based on explicit gender labels and even based on gender-implied first names, behaving differently for male and female players.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A useful exploratory fairness audit whose role-level gender/name swap results are credible in direction, but the headline 'significant' claims are undercut by unstated sampling controls and no released artifacts. the 5 major comments →

arxiv 2508.16065 v1 pith:OYT7AGOD submitted 2025-08-22 cs.CL

Ethical Considerations of Large Language Models in Game Playing

Qingquan Zhang , Yuchen Li , Bo Yuan , Julian Togelius , Georgios N. Yannakakis , Jialin Liu This is my paper
classification cs.CL
keywords large language modelsgender biassocial deduction gamesWerewolfgame AIcounterfactual fairnessname-based gender inferenceLLM ethics
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.

The reading

The paper uses the social deduction game Werewolf as a controlled testbed for a question with practical stakes: do LLM-based agents treat players differently based on gender? It answers yes. Revealing a player's gender changes which players get killed, protected, seen, trusted, and voted for; swapping the genders of other players changes the agent's decisions; and replacing explicit gender labels with strongly gendered first names reproduces the pattern. The effect is not uniform—the Guard and Werewolf are the most gender-sensitive roles, the Seer the most stable—and it has outcome-level consequences, such as female players being killed by Werewolves more often and different win rates by gender. The paper's intended conclusion is that gender bias is not a side issue in LLM game-playing: it degrades fairness and player experience, and removing explicit gender labels is not enough because models infer gender from names.

Core claim

With GLM-3 as the underlying model and a seven-player Werewolf configuration (two Werewolves, three Villagers, one Seer, one Guard), the authors simulate 96 games in which every role is covered by a balanced set of gender configurations. For T1, comparing prompts with and without gender information, they report that in the night (kill/see/protect) and voting scenarios the combined frequency of behavioral change is above 0.5 for nearly every role, and for reliability scoring virtually every state changes when gender is introduced; the Seer is less affected at night. For T2, comparing assign-male and assign-female prompts with hidden-gender behavior, they report that models behave in ways clos

What carries the argument

The load-bearing mechanism is prompt-level counterfactual comparison, quantified by three families of metrics. Freq_s(p) measures how often a player's decision changes when its own stated gender changes from unknown to known; Γ_s measures how similar decisions made without a gender label are to decisions made under an explicitly male or female label; Θ_s measures how often a player's decisions remain unchanged when the genders of all other players are swapped. Each metric is applied to three decision moments: night skills (who to kill, see, or protect), daytime voting, and reliability scoring of other players. The logic of the paper's fairness interpretation is simple: if gender causes behav

Load-bearing premise

The comparisons assume that a single LLM response is a deterministic function of the prompt and game state, so a decision change after a gender swap is caused by gender; if the model is sampled with nonzero temperature and no fixed seed, part of the measured change could be ordinary sampling noise.

What would settle it

Re-run the protocol with temperature 0 and a fixed seed, or repeat each prompt variation many times and average: if gender-swapped decisions change no more often than repeated identical prompts do, and male/female win rates converge, the causal claim fails. A second check is a control set of familiar but gender-neutral names: if those produce the same decision-change frequencies as the seven gendered names, then name-gender association is not the active mechanism.

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

If this is right

  • If the central claim is right, LLM-based agents in Werewolf-like games—at least models comparable to the one tested—will produce systematically unequal outcomes unless prompts and models are audited for gender.
  • Removing explicit gender labels from prompts is not an adequate mitigation, because strongly gendered first names elicit the same biases; effective debiasing has to address latent name–gender associations.
  • The role-dependent sensitivity means fairness evaluation needs to be per-role and per-action rather than a single aggregate score, since Guard and Werewolf can be biased while Seer is comparatively stable.
  • Counterfactual gender-swap metrics of the kind used here can serve as automated fairness tests for social deduction games and similar interactive contexts.
  • Because bias changes target selection and win rates, LLM gender bias in games is not cosmetic—it changes who wins and who is eliminated.

Where Pith is reading between the lines

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

  • One implication the paper leaves open: because no temperature, random seed, or repeated sampling protocol is reported, a share of the measured 'gender-induced' changes may be sample-to-sample stochastic noise. A temperature-0 or repeated-sampling replication would tell how much.
  • The names selected come from U.S. birth records and are strongly gendered there; in other naming cultures the direction and strength of inferred gender could differ, so the specific male/female asymmetries are probably culture-specific even if the general bias mechanism is not.
  • The same name-as-proxy mechanism likely extends beyond games: chat systems, hiring screeners, and negotiation agents that see usernames may show similar gender-conditional behavior even when demographic fields are stripped away.
  • A testable extension: replacing the seven gendered names with matched familiar but gender-neutral names should reduce decision-change frequencies to near zero if gender association is truly the driver; if not, name identity or familiarity, not gender, is doing part of the work.
Share X Bluesky LinkedIn Reddit HN

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

5 major / 6 minor

Summary. The paper studies whether gender information, explicit or implied by first names, changes the behaviour of LLM agents in the social deduction game Werewolf. Using GLM-3 as the agent model, it compares three scenarios—night-time skill use (s1), daytime voting (s2), and reliability/trust assignment (s3)—across prompts that omit, state, or swap gender labels. It reports frequencies of decision changes under gender/name manipulation, interprets values above 0.5 as evidence that gender information significantly affects behaviour, and concludes that LLM agents exhibit role-dependent gender bias, including when gender is conveyed only via names. The paper does not report temperature, seeds, repeated sampling, confidence intervals, or statistical tests, and it does not release code, prompts, or data.

Significance. The topic is timely and the experimental design is conceptually reasonable: counterfactual swaps of gender labels and the use of names as proxies are natural ways to probe bias. If the findings hold, the paper provides a useful demonstration that LLM ethical risks extend to multi-agent social deduction games and that stereotyped first names can trigger discriminatory behaviour. The role-level sensitivity analysis (e.g., Guard and Werewolf being more affected than Seer) could inform model-specific mitigation. However, the paper's empirical foundation is currently incomplete: the absence of stochastic-sampling controls and significance testing means the strong conclusions in Sections 4, 6, and 7 are not yet supported. No machine-checked proofs are claimed, but reproducibility is limited by the lack of released prompts and seeds. The topic and the proxy-name extension give the work a solid base if the statistical issues are addressed.

major comments (5)
  1. [§4.1, Eq. (1); §3.4] The central claim relies on treating each Δs (and similarly Γs and Θs in Tables 1, 2, 4) as a deterministic function of prompt and game state. Section 3.4 does not state a temperature setting, random seed, or repeated-sampling protocol for the main experiments. If GLM-3 is sampled at nonzero temperature, identical prompts at the same game state can yield different actions, so each observed 'change' after gender or name substitution is an upper bound on the causal effect of the manipulation, not an estimate of it. The 70 repetitions in §7.2 do not resolve this unless the same game states are re-sampled under both conditions and the comparisons are paired. Please provide sampling details and repeated paired-run estimates with confidence intervals, or explicitly restrict the claim to temperature-0 decoding.
  2. [§4.2, §6.2, §7.3; §8.3] The text repeatedly uses 'significantly affects behaviour' and 'significant discrimination' without any statistical test or uncertainty quantification. Figures 5, 9–11, and 17 report point frequencies, many around 0.5–0.6, without error bars; Fig. 11's values near 0.12 are presented as 'consistent' although noise alone could produce deviations of this magnitude. A binomial test against a no-gender baseline, a paired permutation test over repeated runs, or at least confidence intervals is required to support the strength of these claims. This is not a presentation issue: the paper itself concedes in §8.3 that measured bias 'may simply result from noise or a small sample size', which is not reconciled with the strong conclusions.
  3. [§3.4; Eq. (1)] The construction of T in Eq. (1) is undefined. Section 3.4 says that the number of male and female players must be a multiple of 48 and that 96 experiments were simulated, but it does not state how gender configurations are enumerated, what T counts, or whether each configuration is played once or multiple times. Without this, Freqs(p) is not reproducible, and the role-level comparisons in Section 4.2 cannot be checked. Please specify the exact mapping from the 96 experiments to the T game states used in Eqs. (1)–(3).
  4. [§5.1, Table 2] The definition of Γ3 is ambiguous. The table states D_p' = 11 − max[|Reliability Score Differences|] and refers to 'a maximum possible reliability score of 11', but the range and scale of the reliability score are never defined. It is unclear what values D_p' can take, whether larger values mean greater similarity, and how scores from different game states are normalized. Since Γ3 feeds the T2/T3 conclusions on reliability assignment, this formulation needs to be specified precisely, including the range of reliability scores and the interpretation of the constant 11.
  5. [§7.2] The seven names are selected by SSA/GenderAPI thresholds >99%, but the paper does not validate that GLM-3 itself consistently maps these names to the intended gender in the Werewolf context. Fig. 16 shows a couple of illustrative examples, but no quantitative per-name association is reported. If the model is uncertain or misassigns some names, the implicit-gender experiment conflates the name manipulation with other lexical associations. Please report per-name gender-recognition rates for the specific model and prompt setting, or use a model-internal validation.
minor comments (6)
  1. [§4.2] Typo: 'sigificantly' should be 'significantly'.
  2. [Table 2] Typo: 'uknown' should be 'unknown'.
  3. [Figs. 10, 11, 14, 19] Figure labels contain 'T otal' (presumably 'Total'). Please correct.
  4. [References] References [18] and [23] appear to duplicate the same paper (Weidinger et al., 'Ethical and social risks of harm from language models'). Please consolidate or correct.
  5. [§5.1] The paragraph beginning 'Eq. 2 aims to capture the similarity...' repeats the same sentence structure multiple times and contains redundant restatements. Please tighten the prose.
  6. [§7.2] The phrase 'retained only those with prediction accuracy scores above 99% according to GenderAPI (genderize)' is vague: it should state whether this is the API's reported confidence, the fraction of correctly classified instances in a validation set, or something else.

Circularity Check

0 steps flagged

No circularity: the paper's claims are empirical measurements with externally sourced controls; no derivation reduces to its own inputs by construction or via a load-bearing self-citation.

full rationale

The paper is an empirical study, not a derivation or prediction from a fitted model. Its central claims—that gender information changes LLM-agent behaviour and that names as gender proxies produce discriminatory behaviour—are operationalized through the metrics in Eqs. (1)–(3) and Tables 1, 2, and 4. These metrics define what counts as a behavioural change or a fairness score, but the paper does not then 'derive' those changes from the metrics as if they were independent facts; it reports measured frequencies. No fitted parameter is later renamed as a prediction. The role-level sensitivity comparisons (e.g., Guard and Werewolf more sensitive than Seer) are descriptive readings of the measured frequencies, not quantities that were used to construct the frequencies. The name-based experiment selects names using an external dataset (U.S. SSA) and an external API (GenderAPI), so the gender association of the names is not imported from the paper's own conclusions; the finding that LLMs react differently to those names is a separate empirical observation. Self-citations appear in the related-work and discussion sections (e.g., refs. [25,26,28,34,35,70]), but none is load-bearing for the main empirical results; the prompt templates and GLM-3 choice are attributed to prior external work [13,41]. The paper even acknowledges in Section 8.3 that 'the bias measured from a given dataset may simply result from noise or a small sample size rather than a genuine systematic effect,' which further confirms that the authors treat the measurements as fallible empirical observations rather than as conclusions forced by definition. No self-definitional reduction, fitted-input-called-prediction, self-citation chain, or ansatz-smuggling via citation is present, so the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

The free parameter list is short because the paper is an empirical observational study rather than a derivation. The key unearned input is the deterministic-output assumption for LLM sampling, which is never stated. The 96-game balancing and 7-name selection are minor experimental choices. No new theoretical entities are introduced.

free parameters (2)
  • Role-gender balancing multiplier = 96 games
    The number 96 is chosen so that every role covers all possible gender identity configurations with equal representation (2 x 2 x 3 x 4). This balancing determines the sample size and is introduced by the authors to make the analysis tractable, but the choice of 96 is not externally motivated.
  • Seven name selections with >99% threshold = Scott, Timothy, Kenneth, Keith, Judith, Mildred, Elizabeth
    The authors selected names with gender assignment percentage above 99% according to SSA and GenderAPI. The threshold and the final seven names are a choice made for the experiment, and the gender ranking of these names is treated as ground truth.
axioms (3)
  • domain assumption LLM outputs for identical prompts are treated as deterministic, so different outputs after gender changes are attributed to the gender change.
    Section 4.1: Freqs(p) = (1/T) Σ Δs(St, p, unknown) compares single runs. If sampling temperature were > 0, output variation alone could change decisions, inflating Δ. The paper never states temperature or seeds.
  • domain assumption The US SSA name-gender distribution is the correct ground truth for name-based gender inference by LLMs.
    Section 7.2: names are selected based on SSA and GenderAPI agreement. The paper assumes GLM-3's name-gender associations match this US-based distribution, which is plausible but unverified for a Chinese model.
  • domain assumption The Werewolf game itself, as instantiated by the authors' prompt pipeline (rules, contextual info, iterative deduction), is a controlled enough environment that behavioral metrics capture ethical/gender bias rather than artifacts of game mechanics.
    Section 3.1 and Figure 2 present the deductive reasoning process. The authors assume the structured environment isolates gender effects, but the pipeline could inject other confounds such as role-specific prompts or dialogue order.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Ethical Considerations of Large Language Models in Game Playing." pith.science (2026). https://pith.science/paper/OYT7AGOD

@misc{pith2026250816065,
  author       = {Pith},
  title        = {Pith review of: Ethical Considerations of Large Language Models in Game Playing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OYT7AGOD}},
  note         = {Machine review of arXiv:2508.16065}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large language models (LLMs) have demonstrated tremendous potential in game playing, while little attention has been paid to their ethical implications in those contexts. This work investigates and analyses the ethical considerations of applying LLMs in game playing, using Werewolf, also known as Mafia, as a case study. Gender bias, which affects game fairness and player experience, has been observed from the behaviour of LLMs. Some roles, such as the Guard and Werewolf, are more sensitive than others to gender information, presented as a higher degree of behavioural change. We further examine scenarios in which gender information is implicitly conveyed through names, revealing that LLMs still exhibit discriminatory tendencies even in the absence of explicit gender labels. This research showcases the importance of developing fair and ethical LLMs. Beyond our research findings, we discuss the challenges and opportunities that lie ahead in this field, emphasising the need for diving deeper into the ethical implications of LLMs in gaming and other interactive domains.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Even More Deception: Objective Misalignment in Mixed-Motive LLM Multi-Agent Systems

    cs.AI 2026-07 conditional novelty 6.0

    Changing one LLM agent's secret objective in Werewolf lowers its team's win rate and changes its reasoning, while its public chat stays deceptively normal.

Reference graph

Works this paper leans on

69 extracted references · 54 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    Deepseek llm: Scaling open-source language models with longtermism

    Bi X, Chen D, Chen G, Chen S, Dai D, Deng C, Ding H, Dong K, Du Q, Fu Z, others . Deepseek llm: Scaling open-source language models with longtermism. arXiv preprint arXiv:2401.02954, 2024

    Pith/arXiv arXiv 2024

  2. [2]

    GPT-4 technical report, 2024

    OpenAI . GPT-4 technical report, 2024

    work page 2024

  3. [3]

    ChatGLM: A family of large language models from glm-130b to glm-4 all tools

    GLM T, Zeng A, Xu B, Wang B, Zhang C, Yin D, Zhang D, Rojas D, Feng G, Zhao H, others . ChatGLM: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Large language models encode clinical knowledge

    Singhal K, Azizi S, Tu T, Mahdavi S S, Wei J, Chung H W, Scales N, Tanwani A, Cole-Lewis H, Pfohl S, others . Large language models encode clinical knowledge. Nature, 2023, 620(7972): 172–180

    work page 2023

  5. [5]

    Causal reasoning and large language models: Opening a new frontier for causality

    Kıcıman E, Ness R, Sharma A, Tan C. Causal reasoning and large language models: Opening a new frontier for causality. arXiv preprint arXiv:2305.00050, 2023

    Pith/arXiv arXiv 2023

  6. [6]

    Large lan- guage models make sample-efficient recommender systems

    Lin J, Dai X, Shan R, Chen B, Tang R, Yu Y , Zhang W. Large lan- guage models make sample-efficient recommender systems. Frontiers of Computer Science, 2025, 19(4): 194328

    work page 2025

  7. [7]

    Large language models and games: A survey and roadmap

    Gallotta R, Todd G, Zammit M, Earle S, Liapis A, Togelius J, Yan- nakakis G N. Large language models and games: A survey and roadmap. IEEE Transactions on Games, 2024, 1–18

    work page 2024

  8. [8]

    A survey of large language models

    Zhao W X, Zhou K, Li J, Tang T, Wang X, Hou Y , Min Y , Zhang B, Zhang J, Dong Z, others . A survey of large language models. arXiv preprint arXiv:2303.18223, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Science in the age of large language models

    Birhane A, Kasirzadeh A, Leslie D, Wachter S. Science in the age of large language models. Nature Reviews Physics, 2023, 5(5): 277–280

    work page 2023

  10. [10]

    A survey on large language model-based game agents

    Hu S, Huang T, Ilhan F, Tekin S, Liu G, Kompella R, Liu L. A survey on large language model-based game agents. arXiv preprint arXiv:2404.02039, 2024

    Pith/arXiv arXiv 2024

  11. [11]

    Exploring large language models for communication games: An empirical study on Werewolf

    Xu Y , Wang S, Li P, Luo F, Wang X, Liu W, Liu Y . Exploring large language models for communication games: An empirical study on Werewolf. arXiv preprint arXiv:2309.04658, 2023

    Pith/arXiv arXiv 2023

  12. [12]

    Werewolf Among Us: Multimodal resources for modeling persuasion behaviors in social deduction games

    Lai B, Zhang H, Liu M, Pariani A, Ryan F, Jia W, Hayati S A, Rehg J, Yang D. Werewolf Among Us: Multimodal resources for modeling persuasion behaviors in social deduction games. 2023, 6570–6588 Front. Comput. Sci., 2025, 0(0): 1–19

    work page 2023

  13. [13]

    Helmsman of the masses? evaluate the opinion lead- ership of large language models in the Werewolf game

    Du S, Zhang X. Helmsman of the masses? evaluate the opinion lead- ership of large language models in the Werewolf game. In: First Con- ference on Language Modeling. 2024

    work page 2024

  14. [14]

    LLM-based agent society investigation: Collaboration and confrontation in Avalon gameplay

    Lan Y , Hu Z, Wang L, Wang Y , Ye D, Zhao P, Lim E P, Xiong H, Wang H. LLM-based agent society investigation: Collaboration and confrontation in Avalon gameplay. arXiv preprint arXiv:2310.14985, 2023

    Pith/arXiv arXiv 2023

  15. [15]

    How far are we on the decision-making of LLMs? evaluating LLMs’ gaming ability in multi-agent environments

    Huang t J, Li E J, Lam M H, Liang T, Wang W, Yuan Y , Jiao W, Wang X, Tu Z, Lyu M R. How far are we on the decision-making of LLMs? evaluating LLMs’ gaming ability in multi-agent environments. CoRR, 2024, abs/2403.11807

    Pith/arXiv arXiv 2024

  16. [16]

    A survey on multilingual large language models: Corpora, alignment, and bias

    Xu Y , Hu L, Zhao J, Qiu Z, Xu K, Ye Y , Gu H. A survey on multilingual large language models: Corpora, alignment, and bias. Frontiers of Computer Science, 2025, 19(11): 1911362

    work page 2025

  17. [17]

    TrustLLM: Trustworthiness in large language models

    Huang Y , Sun L, Wang H, Wu S, Zhang Q, Li Y , Gao C, Huang Y , Lyu W, Zhang Y , others . TrustLLM: Trustworthiness in large language models. arXiv preprint arXiv:2401.05561, 2024

    Pith/arXiv arXiv 2024

  18. [19]

    A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly

    Yao Y , Duan J, Xu K, Cai Y , Sun Z, Zhang Y . A survey on large language model (LLM) security and privacy: The good, the bad, and the ugly. High-Confidence Computing, 2024, 4(2): 100211

    work page 2024

  19. [20]

    Fairness in large language models: A tax- onomic survey

    Chu Z, Wang Z, Zhang W. Fairness in large language models: A tax- onomic survey. ACM SIGKDD explorations newsletter, 2024, 26(1): 34–48

    work page 2024

  20. [21]

    An overview of artificial intelli- gence ethics

    Huang C, Zhang Z, Mao B, Yao X. An overview of artificial intelli- gence ethics. IEEE Transactions on Artificial Intelligence, 2023, 4(4): 799–819

    work page 2023

  21. [22]

    Per- turbation augmentation for fairer NLP

    Qian R, Ross C, Fernandes J, Smith E M, Kiela D, Williams A. Per- turbation augmentation for fairer NLP. In: Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022, 9496–9521

    work page 2022

  22. [23]

    Ethical and social risks of harm from language models

    Weidinger L, Mellor J, Rauh M, Gri ffin C, Uesato J, Huang P S, Cheng M, Glaese M, Balle B, Kasirzadeh A, others . Ethical and social risks of harm from language models. arXiv preprint arXiv:2112.04359, 2021

    Pith/arXiv arXiv 2021

  23. [24]

    Improving LLM group fairness on tabular data via in-context learning

    Cherepanova V , Lee C J, Akpinar N J, Fogliato R, Bertran M A, Kearns M, Zou J. Improving LLM group fairness on tabular data via in-context learning. In: Neurips Safe Generative AI Workshop 2024

    work page 2024

  24. [25]

    Fairer machine learn- ing through multi-objective evolutionary learning

    Zhang Q, Liu J, Zhang Z, Wen J, Mao B, Yao X. Fairer machine learn- ing through multi-objective evolutionary learning. In: International conference on artificial neural networks. 2021, 111–123

    work page 2021

  25. [26]

    Mitigating unfairness via evolutionary multiobjective ensemble learning

    Zhang Q, Liu J, Zhang Z, Wen J, Mao B, Yao X. Mitigating unfairness via evolutionary multiobjective ensemble learning. IEEE Transactions on Evolutionary Computation, 2023, 27(4): 848–862

    work page 2023

  26. [27]

    Fairness is essential for robustness: fair adversarial training by identifying and augmenting hard examples

    Mou N, Yue X, Zhao L, Wang Q. Fairness is essential for robustness: fair adversarial training by identifying and augmenting hard examples. Frontiers of Computer Science, 2025, 19(3): 193803

    work page 2025

  27. [28]

    Fairer machine learning through the hybrid of multi-objective evolutionary learning and adversarial learn- ing

    Gui S, Zhang Q, Huang C, Yuan B. Fairer machine learning through the hybrid of multi-objective evolutionary learning and adversarial learn- ing. In: 2023 International Joint Conference on Neural Networks (IJCNN). 2023, 1–9

    work page 2023

  28. [29]

    Ethical considerations and potential risks in the deployment of large language models in diverse societal contexts

    Liyanage U P, Ranaweera N D. Ethical considerations and potential risks in the deployment of large language models in diverse societal contexts. Journal of Computational Social Dynamics, 2023, 8(11): 15– 25

    work page 2023

  29. [30]

    Consent-GPT: is it eth- ical to delegate procedural consent to conversational AI? Journal of Medical Ethics, 2024, 50(2): 77–83

    Allen J W, Earp B D, Koplin J, Wilkinson D. Consent-GPT: is it eth- ical to delegate procedural consent to conversational AI? Journal of Medical Ethics, 2024, 50(2): 77–83

    work page 2024

  30. [31]

    Determinants of LLM-assisted decision-making

    Eigner E, H ¨andler T. Determinants of LLM-assisted decision-making. arXiv preprint arXiv:2402.17385, 2024

    Pith/arXiv arXiv 2024

  31. [32]

    RLupus: Cooperation through emer- gent communication in the Werewolf social deduction game

    Brandizzi N, Grossi D, Iocchi L. RLupus: Cooperation through emer- gent communication in the Werewolf social deduction game. Intelli- genza Artificiale, 2022, 15(2): 55–70

    work page 2022

  32. [33]

    The ethics of AI in games

    Melhart D, Togelius J, Mikkelsen B, Holmgård C, Yannakakis G N. The ethics of AI in games. IEEE Transactions on Affective Computing, 2024, 15(1): 79–92

    work page 2024

  33. [34]

    Exploring Accuracy-Fairness Trade-off in Large Language Models

    Zhang Q, Duan Q, Yuan B, Shi Y , Liu J. Exploring accuracy-fairness trade-off in large language models. arXiv preprint arXiv:2411.14500, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  34. [35]

    Fair- erML: An extensible platform for analysing, visualising, and mitigat- ing biases in machine learning [application notes]

    Yuan B, Gui S, Zhang Q, Wang Z, Wen J, Mao B, Liu J, Yao X. Fair- erML: An extensible platform for analysing, visualising, and mitigat- ing biases in machine learning [application notes]. IEEE Computa- tional Intelligence Magazine, 2024, 19(2): 129–141

    work page 2024

  35. [36]

    The risks of us- ing ChatGPT to obtain common safety-related information and advice

    Oviedo-Trespalacios O, Peden A E, Cole-Hunter T, Costantini A, Haghani M, Rod J, Kelly S, Torkamaan H, Tariq A, David Albert New- ton J, Gallagher T, Steinert S, Filtness A J, Reniers G. The risks of us- ing ChatGPT to obtain common safety-related information and advice. Safety Science, 2023, 167: 106244

    work page 2023

  36. [37]

    Artificial intelligence and games

    Yannakakis G N, Togelius J. Artificial intelligence and games. vol- ume 2. Springer, 2018

    work page 2018

  37. [38]

    Deep reinforcement learning for general video game AI

    Torrado R R, Bontrager P, Togelius J, Liu J, Perez-Liebana D. Deep reinforcement learning for general video game AI. In: 2018 IEEE Conference on Computational Intelligence and Games (CIG). 2018, 1–8

    work page 2018

  38. [39]

    Why do women pretend to be men? female gender swapping in online games

    Zhou L, Han N, Xu Z, Brian C, Hussain S. Why do women pretend to be men? female gender swapping in online games. Frontiers in psychology, 2022, 13: 810954

    work page 2022

  39. [40]

    Gender bias in video game dialogue

    Rennick S, Clinton M, Ioannidou E, Oh L, Clooney C, Healy E, Roberts S G. Gender bias in video game dialogue. Royal Society Open Science, 2023, 10(5): 221095

    work page 2023

  40. [41]

    Language agents with reinforce- ment learning for strategic play in the werewolf game

    Xu Z, Yu C, Fang F, Wang Y , Wu Y . Language agents with reinforce- ment learning for strategic play in the werewolf game. In: Forty-first International Conference on Machine Learning. 2023

    work page 2023

  41. [42]

    Overview of AIWolfDial 2019 shared task: Contest of automatic dialog agents to play the Werewolf game through conversations

    Kano Y , Aranha C, Inaba M, Toriumi F, Osawa H, Katagami D, Otsuki T, Tsunoda I, Nagayama S, Tellols D, Sugawara Y , Nakata Y . Overview of AIWolfDial 2019 shared task: Contest of automatic dialog agents to play the Werewolf game through conversations. In: Kano Y , Aranha C, Inaba M, Toriumi F, Osawa H, Katagami D, Otsuki T, eds, Proceedings of the 1st In...

    work page 2019

  42. [43]

    Enhance reason- ing for large language models in the game Werewolf

    Wu S, Zhu L, Yang T, Xu S, Fu Q, Wei Y , Fu H. Enhance reason- ing for large language models in the game Werewolf. arXiv preprint arXiv:2402.02330, 2024 Front. Comput. Sci., 2025, 0(0): 1–19

    Pith/arXiv arXiv 2024

  43. [44]

    Enhancing dialogue generation in Werewolf game through situation analysis and persuasion strategies

    Qi Z, Inaba M. Enhancing dialogue generation in Werewolf game through situation analysis and persuasion strategies. In: Proceedings of the 2nd International AIWolfDial Workshop. 2024, 30–39

    work page 2024

  44. [45]

    Enhancing consistency of Werewolf AI through dialogue summarization and persona information

    Tanaka Y , Kaneko T, Onozeki H, Ezure N, Uehara R, Qi Z, Higuchi T, Asahara R, Inaba M. Enhancing consistency of Werewolf AI through dialogue summarization and persona information. In: Kano Y , ed, Pro- ceedings of the 2nd International AIWolfDial Workshop. September 2024, 48–57

    work page 2024

  45. [46]

    A theory of social comparison processes

    Festinger L. A theory of social comparison processes. Human rela- tions, 1954, 7(2): 117–140

    work page 1954

  46. [47]

    Counterfactual fairness

    Kusner M J, Loftus J, Russell C, Silva R. Counterfactual fairness. Advances in neural information processing systems, 2017, 30

    work page 2017

  47. [48]

    A roadmap of explainable artificial intelli- gence: Explain to whom, when, what and how? ACM Trans

    Wang Z, Huang C, Yao X. A roadmap of explainable artificial intelli- gence: Explain to whom, when, what and how? ACM Trans. Auton. Adapt. Syst., 2024, 19(4)

    work page 2024

  48. [49]

    War and peace (WarAgent): Large language model-based multi-agent simula- tion of world wars

    Hua W, Fan L, Li L, Mei K, Ji J, Ge Y , Hemphill L, Zhang Y . War and peace (WarAgent): Large language model-based multi-agent simula- tion of world wars. arXiv preprint arXiv:2311.17227, 2023

    Pith/arXiv arXiv 2023

  49. [50]

    It’s all in the name: Mitigating gender bias with name-based counterfactual data substitu- tion

    Gonen H, Cotterell R, Maudslay R H, Teufel S. It’s all in the name: Mitigating gender bias with name-based counterfactual data substitu- tion. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Confer- ence on Natural Language Processing (EMNLP-IJCNLP). 2019

    work page 2019

  50. [51]

    What’s in a name?–gender classification of names with character based machine learning models

    Hu Y , Hu C, Tran T, Kasturi T, Joseph E, Gillingham M. What’s in a name?–gender classification of names with character based machine learning models. Data Mining and Knowledge Discovery, 2021, 35(4): 1537–1563

    work page 2021

  51. [52]

    Beyond binary gender labels: Revealing gender bias in llms through gender- neutral name predictions

    You Z, Lee H, Mishra S, Jeoung S, Mishra A, Kim J, Diesner J. Beyond binary gender labels: Revealing gender bias in llms through gender- neutral name predictions. In: Proceedings of the 5th Workshop on Gender Bias in Natural Language Processing (GeBNLP). 2024, 255– 268

    work page 2024

  52. [53]

    First-person fair- ness in chatbots

    Eloundou T, Beutel A, Robinson D G, Gu-Lemberg K, Brakman A L, Mishkin P, Shah M, Heidecke J, Weng L, Kalai A T. First-person fair- ness in chatbots. In: The Twelfth International Conference on Learning Representations. 2025

    work page 2025

  53. [54]

    National data on the relative frequency of given names in the population of u.s

    Administration S S. National data on the relative frequency of given names in the population of u.s. births where the individual has a so- cial security number (tabulated based on social security records as of march 3, 2019). In: The Twelfth International Conference on Learning Representations. 2018

    work page 2019

  54. [55]

    AI Gender Bias, Disparities, and Fairness: Does Training Data Matter?

    Latif E, Zhai X, Liu L. AI gender bias, disparities, and fairness: Does training data matter? arXiv preprint arXiv:2312.10833, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  55. [56]

    The e ffect of gender stereotypes on artificial intelligence recommendations

    Ahn J, Kim J, Sung Y . The e ffect of gender stereotypes on artificial intelligence recommendations. Journal of Business Research, 2022, 141: 50–59

    work page 2022

  56. [57]

    Improving gender fairness of pre- trained language models without catastrophic forgetting

    Fatemi Z, Xing C, Liu W, Xiong C. Improving gender fairness of pre- trained language models without catastrophic forgetting. In: Rogers A, Boyd-Graber J, Okazaki N, eds, Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume 2: Short Papers). July 2023, 1249–1262

    work page 2023

  57. [58]

    Mixup-based unified framework to over- come gender bias resurgence

    Yu L, Mao Y , Wu J, Zhou F. Mixup-based unified framework to over- come gender bias resurgence. In: Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’23. 2023, 1755–1759

    work page 2023

  58. [59]

    Auto-debias: Debiasing masked language models with automated biased prompts

    Guo Y , Yang Y , Abbasi A. Auto-debias: Debiasing masked language models with automated biased prompts. In: Muresan S, Nakov P, Villavicencio A, eds, Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers). May 2022, 1012–1023

    work page 2022

  59. [60]

    Learnware: on the future of machine learning

    Zhou Z H. Learnware: on the future of machine learning. Frontiers of Computer Science, 2016, 10(4): 589–590

    work page 2016

  60. [61]

    Explicitly unbiased large language models still form biased associations

    Bai X, Wang A, Sucholutsky I, Gri ffiths T L. Explicitly unbiased large language models still form biased associations. Proceedings of the National Academy of Sciences, 2025, 122(8): e2416228122

    work page 2025

  61. [62]

    The global landscape of AI ethics guide- lines

    Jobin A, Ienca M, Vayena E. The global landscape of AI ethics guide- lines. Nature machine intelligence, 2019, 1(9): 389–399

    work page 2019

  62. [63]

    The ethics of AI ethics: An evaluation of guidelines

    Hagendor ff T. The ethics of AI ethics: An evaluation of guidelines. Minds and machines, 2020, 30(1): 99–120

    work page 2020

  63. [64]

    Semantics derived automati- cally from language corpora contain human-like biases

    Caliskan A, Bryson J J, Narayanan A. Semantics derived automati- cally from language corpora contain human-like biases. Science, 2017, 356(6334): 183–186

    work page 2017

  64. [65]

    Bias in data-driven artificial intelligence systems—an introductory survey

    Ntoutsi E, Fafalios P, Gadiraju U, Iosifidis V , Nejdl W, Vidal M E, Ruggieri S, Turini F, Papadopoulos S, Krasanakis E, others . Bias in data-driven artificial intelligence systems—an introductory survey. Wi- ley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 2020, 10(3): e1356

    work page 2020

  65. [66]

    How well do LLMs represent values across cultures? empirical analysis of LLMs re- sponses based on hofstede cultural dimensions

    Kharchenko J, Roosta T, Chadha A, Shah C. How well do LLMs represent values across cultures? empirical analysis of LLMs re- sponses based on hofstede cultural dimensions. arXiv preprint arXiv:2406.14805, 2024

    Pith/arXiv arXiv 2024

  66. [67]

    AI pitfalls and what not to do: mitigating bias in AI

    Gichoya J W, Thomas K, Celi L A, Safdar N, Banerjee I, Banja J D, Seyyed-Kalantari L, Trivedi H, Purkayastha S. AI pitfalls and what not to do: mitigating bias in AI. The British Journal of Radiology, 2023, 96(1150): 20230023

    work page 2023

  67. [68]

    Embracing LLM feedback: the role of feed- back providers and provider information for feedback effectiveness

    Ruwe T, Mayweg-Paus E. Embracing LLM feedback: the role of feed- back providers and provider information for feedback effectiveness. In: Frontiers in Education. 2024, 1461362

    work page 2024

  68. [69]

    Prompt injection attack against llm-integrated applications

    Liu Y , Deng G, Li Y , Wang K, Wang Z, Wang X, Zhang T, Liu Y , Wang H, Zheng Y , others . Prompt injection attack against llm-integrated applications. arXiv preprint arXiv:2306.05499, 2023

    Pith/arXiv arXiv 2023

  69. [70]

    Fairness-aware multiobjective evolution- ary learning

    Zhang Q, Liu J, Yao X. Fairness-aware multiobjective evolution- ary learning. IEEE Transactions on Evolutionary Computation, 2024, doi:10.1109/TEVC.2024.3430824, 1–14

    work page arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.