Pith. sign in

REVIEW 5 major objections 6 minor 20 references

Large Language Models as Mirrors of Societal Moral Standards

T0 review · 5 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper finds that pre-trained language models, probed on WVS and PEW moral topics, capture some cultural value differences but align only weakly with survey-based moral standards; the multilingual BLOOMZ-560M performs best yet still…

desk verdict A replication that confirms known weak alignment between open PLMs and cross-national moral surveys, but the quantitative claims about BLOOMZ rest on an under-specified token-scoring step that needs fixing before effect sizes are taken seriously. read the letter →

arxiv 2412.00956 v1 pith:ZLICEKZZ submitted 2024-12-01 cs.AI cs.CLcs.SC

Evi Papadopoulou , Hadi Mohammadi , Ayoub Bagheri This is my paper · ORCID
classification cs.AIcs.CLcs.SC
keywords moralnormsculturalvalueslargelanguagemodelsWorldSurveyPEWGlobalAttitudescross-culturalalignmentlog-probabilityprobingBLOOMZ
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 pre-trained language models encode the moral norms of specific countries well enough to mirror human survey responses. Using moral-topic questions from the World Values Survey and the PEW Global Attitudes Survey as ground truth, the authors probe GPT-2, OPT, BLOOMZ-560M, and Qwen2 with English prompts of the form 'In {country}, {topic} is {moral_judgment}.' They conclude that the models do capture some cultural value differences, but correlations with survey means are weak and often negative, and models tend to judge most topics as acceptable regardless of country. The best performer, BLOOMZ-560M, reaches averaged correlations of 0.25 and 0.29 on WVS and 0.16 and 0.11 on PEW, which the authors still consider far from a moderate reflection of human opinion. The study thus challenges the idea that current pre-trained language models can serve as reliable mirrors of societal moral standards.

What carries the argument

The central object is the moral score, defined as the difference between the log-probability of a morally charged token and the log-probability of its opposing token, averaged over five pairs of moral judgments such as (always justifiable, never justifiable), (right, wrong), and (ethical, unethical). Concretely, $\text{moral\_score} = \frac{1}{5}\sum_i (\log p_{\text{moral},i} - \log p_{\text{nonmoral},i})$ computed from the next-token distribution after an English template prompt. This score is the bridge between models and surveys: it is correlated with the averaged normalized survey responses per country-topic pair, and the correlation coefficients are the paper's main evidence. The choice of token pair, not the prompt phrasing, drives most of the variation in results.

What would settle it

If correlations with survey means remain at the same magnitude when the morally charged token pairs are replaced by non-moral positive/negative word pairs such as 'good' and 'bad' or 'pleasant' and 'unpleasant', then the moral score is tracking general valence rather than country-specific moral knowledge, falsifying the paper's interpretation of its correlations.

Watch

Extended reading notes

Core claim

The central claim is that pre-trained language models do not reliably encode cross-cultural moral norms. For each of 19 WVS moral topics across 55 countries and 8 PEW topics across 39 countries, the authors compute a moral score from the model's next-token log probabilities and compare it with the country-topic mean from the surveys. All four model families, GPT-2, OPT, BLOOMZ-560M, and Qwen2, produce correlations that are weak, frequently negative, and highly sensitive to which moral token pair is used. BLOOMZ-560M, the multilingual model fine-tuned for zero-shot tasks, shows the closest alignment, with averaged correlations around 0.25 to 0.29 for WVS and 0.11 to 0.16 for PEW, yet the paper concludes that even this model failed to reflect human opinions even to a moderate degree. The discovery is more a negative result and replication check: prior claims about cultural moral knowledge in language models do not survive this broader, multi-model comparison.

Load-bearing premise

The load-bearing premise is that a model's log-probability gap between a morally charged token and its opposing token, measured in an English prompt, reflects the model's moral stance on that country-topic pair rather than token frequency, English-language priors, or a general positivity bias.

Editorial extensions

If this is right

  • Value-sensitive applications such as content moderation or cross-cultural chatbots should not assume that models align with local moral standards; the weak and often negative correlations indicate systematic mismatch.
  • The relative success of BLOOMZ-560M suggests that multilingual training plus zero-shot fine-tuning is a more promising direction for cultural alignment than English-language scale alone.
  • Because moral token choice changes results more than prompt phrasing, any claim about a model's moral knowledge must be qualified by the exact probe vocabulary used.
  • The tendency of all models to produce positive moral scores for most topics implies a positivity or acceptability bias that will require explicit mitigation before models can mirror survey-style moral judgments.
  • The results temper earlier findings that English pre-trained models capture cultural moral norms, extending the replication to multilingual models and two large survey datasets.

Reading between the lines

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

  • Not tested in the paper, the moral score may be measuring general valence rather than culturally specific morality; a direct test would replace moral adjectives with matched non-moral adjectives and check whether country correlations persist.
  • Since all prompts are in English, the weak performance of multilingual models could reflect an English-language prior rather than an absence of cultural knowledge; probing in each country's dominant language is a testable extension.
  • Averaging survey responses to country means discards within-country disagreement; models could match some population segments while failing to match the mean, which would change the interpretation of low correlations.
  • The paper's own discussion notes that Spearman correlations, newer models, or larger parameter variants might alter results, so the negative conclusion is contingent on this specific probe setup.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper probes four autoregressive language models (GPT-2 variants, OPT variants, Qwen2-0.5B, BLOOMZ-560M) for cross-cultural moral knowledge by scoring English prompts of the form 'In {country} {topic} is {moral_judgment}' and 'People in {country} believe {topic} is {moral_judgment}.', computing a 'moral score' as the difference in log-probabilities between morally charged and non-morally charged token pairs, and correlating these scores with country-topic means from the World Values Survey Wave 7 and the PEW 2013 Global Attitudes Survey. The main finding is that most models show weak or negative correlations with survey data, while BLOOMZ-560M shows the strongest positive correlations (e.g., 0.25 and 0.29 on WVS averaged token pairs), yet still falls short of moderate alignment. The paper concludes that current PLMs capture some but not reliable cross-cultural moral variation and that token-pair choice matters more than prompt phrasing.

Significance. If the measurement were validated, the paper would provide a useful replication and extension of prior work on cultural moral norms in LLMs, with an explicit comparison of monolingual and multilingual models across two large survey datasets. The study is transparent about model sizes and uses publicly available benchmarks, and the conclusion that even the best model only weakly aligns with survey ground truth is a sobering and potentially valuable result for the growing literature on cultural AI alignment. However, the paper's central quantitative claim rests on a moral-score computation that is ambiguously specified for multi-token pairs and that does not control for token-frequency or positivity baselines; without resolving these issues, the reported correlations (including the headline BLOOMZ result) are not interpretable as evidence about cultural moral knowledge. The paper also makes several non-defended preprocessing choices (replacing WVS missing values with 0, coding PEW non-responses as morally unacceptable) that can systematically distort the survey ground truth.

major comments (5)
  1. [§4.2 (Method for Moral Score Calculation)] The definition of the moral score is ambiguous for multi-token moral judgments. The text states that 'the logit corresponding to the word appearing after the prompt is calculated,' but pairs 1, 3, and 4 are 'always justifiable' vs. 'never justifiable', 'morally good' vs. 'morally bad', and 'ethically right' vs. 'ethically wrong'—all multi-word phrases. If only the first token is scored, pair3 and pair4 have identical first tokens in both alternatives, making the score identically zero and impossible to yield the nonzero correlations reported in Tables 1–10. If the full phrase is scored as a sequence (e.g., by summing log-probabilities), the procedure is not described. The paper must specify exactly how multi-token pairs are scored, provide the corresponding formulas, and ideally release the scoring code. Until this is resolved, the central correlations, including the BLOOMZ results in Tables 9 and 10, cannot be interpreted.
  2. [§4.2 (no baseline control)] The moral score is computed as log P(moral token) − log P(nonmoral token) with no subtraction of a country-agnostic or topic-agnostic baseline. For a pair like (right, wrong), the difference can be dominated by the marginal frequency of the two words in English or by a general positivity bias, independent of the country or topic in the prompt. Correlating this unadjusted score with survey means across country-topic pairs can therefore produce spurious alignment driven by topic-level priors or overall response tendencies. The authors should show that a control condition—for example, the same token pairs with the country and topic removed or replaced by a neutral word—does not reproduce the reported correlations. Without such a baseline, the claim that BLOOMZ 'captures certain cultural value differences' is not established.
  3. [§4.1 (WVS preprocessing)] Replacing missing WVS responses (values −1, −2, −4, −5) with 0 on the original 1–10 scale is a non-neutral choice. Because 0 lies outside the valid response range and is lower than the minimum valid response (1), countries with higher nonresponse rates will have artificially deflated mean moral scores. This systematically biases the country-topic ground truth and can inflate or deflate correlations with model scores depending on the missingness pattern. The authors should either exclude nonresponses before averaging, use survey weights, or justify why a 0-imputation is unbiased; a sensitivity analysis with listwise deletion would be a minimal check.
  4. [§4.1 (PEW preprocessing)] For the PEW data, the paper assigns 'Depends on situation', 'Refused', and 'Don't know' responses the same code (−1) as 'morally unacceptable'. This conflates non-attitudes or missingness with a strong negative moral judgment and will pull country-topic means downward for countries with higher nonresponse or 'depends' rates. The resulting survey scores do not faithfully represent respondent opinions, undermining the ground truth against which model scores are correlated. These responses should be treated as missing (excluded) or handled in a way that does not equate them with 'morally unacceptable'.
  5. [§5.2 and Tables 9–10 (BLOOMZ claim)] The paper's headline positive result—BLOOMZ-560M's averaged correlations of 0.25/0.29 on WVS and 0.16/0.11 on PEW—is computed by averaging over five token pairs that include the ambiguous multi-token pairs 1, 3, and 4. Given the unresolved scoring ambiguity and the lack of a baseline control, these numbers cannot currently be taken as evidence that BLOOMZ approximates human moral opinions even weakly. The authors need to rerun the analysis with an unambiguous scoring rule and a baseline subtraction, and then report whether the ordering of models and the significance of the positive correlations survive.
minor comments (6)
  1. [§3.1 vs. §4.1] The dataset is described as WVS Wave 7 in Section 3.1, while Section 4.1 refers to 'Version 5' of the WVS data; the reference list also cites Round Six. Please clarify which wave/version was actually used and correct the reference.
  2. [§4.1 (PEW coding)] The description of the PEW response coding is easy to misread: it states that 'Not a moral issue' is assigned 0 and non-responses −1, but the previous sentence says responses were normalized between −1 and 1. Clarify whether 'morally unacceptable' itself is −1 and whether the normalization is applied after this coding.
  3. [§4.2 and Tables 1–10] The significance stars are not tied to exact p-values or confidence intervals, and some tables have an empty p-value column. Reporting exact p-values or confidence intervals would make the correlation results more interpretable, especially given the large number of comparisons.
  4. [§5.1] The text states that 'the strongest negative correlation appears with pair5 in the 'in' mode' for GPT-2 base on WVS, but Table 1 shows pair1 'in' at −0.39 and pair3 'people' at −0.28; the stated 'strongest' value does not match the table. Please reconcile the narrative with the table entries.
  5. [§4.1 (PEW topics)] The paper lists sample topics like 'using contraceptives' and 'getting a divorce' but does not provide the full list of the eight topics Q84A–Q84H or the exact prompt phrases used for each topic. Include the full topic list and the resulting prompt examples to make the experiments reproducible.
  6. [General reproducibility] No code or processed data is released, and the Hugging Face model versions are not pinned with commit hashes. Given the dependence on exact tokenization and normalization procedures, releasing the scoring code and processed survey means is important for verifying the results.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the model scores are independent of the survey ground truth they are correlated against.

full rationale

The paper's central claim—that LLM moral scores correlate only weakly with WVS and PEW survey means—rests on an external benchmark comparison. The language_model_score in Section 4.2 is defined as moral_logprob minus nonmoral_logprob and averaged over five token pairs, with no parameter fitted to the survey data. The survey means are used only as ground truth for Pearson correlation after separate normalization, and the normalization is affine, so it does not change the correlation. The prompts are derived from survey topics, but this is standard benchmarking practice and does not make the model outputs equal to the survey values by construction. The cited prior work (Ramezani and Xu, 2023; Arora et al., 2022) is by other authors, so there is no self-citation chain. Ambiguities about scoring multiword token pairs and the absence of a frequency baseline are validity or correctness concerns, not instances of circularity, because the model scores are not constructed from the target survey responses. Therefore no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The evaluation rests on the proxy validity of the moral score and on treating survey averages as ground truth. No fitted parameters or new entities are introduced; the main unexamined choices are the missing-value recoding and the English-only probing of multilingual models.

assumptions (5)
  • domain assumption Country-topic means from WVS and PEW surveys are a valid ground truth for societal moral norms.
    The entire evaluation treats these survey aggregates as the target values the models should mirror.
  • domain assumption The difference of log-probabilities of moral vs nonmoral token completions is a valid measure of a model's moral stance.
    Section 4.2 defines moral_score as moral_logprob minus nonmoral_logprob and uses it as the model's opinion.
  • domain assumption English prompts can elicit cross-cultural moral knowledge from multilingual models.
    All prompts are in English, including for BLOOMZ and Qwen2, with no translation or language-specific probing.
  • ad hoc to paper Replacing missing WVS responses with 0 does not systematically bias country-topic means.
    Section 4.1 recodes 'Don't know', 'No answer', 'Not asked', and 'Missing' to 0, but the paper supplies no justification that this is unbiased.
  • domain assumption The 100 respondents per country in the PEW dataset are representative.
    Section 3.2 states the dataset has 100 respondents per country; the paper treats these as representative of national opinion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Language Models as Mirrors of Societal Moral Standards." pith.science (2026). https://pith.science/paper/ZLICEKZZ

@misc{pith2026241200956,
  author       = {Pith},
  title        = {Pith review of: Large Language Models as Mirrors of Societal Moral Standards},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZLICEKZZ}},
  note         = {Machine review of arXiv:2412.00956}
}
read the original abstract

Prior research has demonstrated that language models can, to a limited extent, represent moral norms in a variety of cultural contexts. This research aims to replicate these findings and further explore their validity, concentrating on issues like 'homosexuality' and 'divorce'. This study evaluates the effectiveness of these models using information from two surveys, the WVS and the PEW, that encompass moral perspectives from over 40 countries. The results show that biases exist in both monolingual and multilingual models, and they typically fall short of accurately capturing the moral intricacies of diverse cultures. However, the BLOOM model shows the best performance, exhibiting some positive correlations, but still does not achieve a comprehensive moral understanding. This research underscores the limitations of current PLMs in processing cross-cultural differences in values and highlights the importance of developing culturally aware AI systems that better align with universal human values.

Figures

Figures reproduced from arXiv: 2412.00956 by the authors.

Figure 3
Figure 3. Distribution of normalized answer values for [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Spread of responses across the moral topics [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 6
Figure 6. Distribution of normalized moral scores from [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Distribution of normalized moral scores from [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 7
Figure 7. Figure 7: Distribution of normalized moral scores from [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Distribution of normalized moral scores from [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 11
Figure 11. Figure 11: Distribution of normalized moral scores from [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 12
Figure 12. Figure 12: Distribution of normalized moral scores from [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 2 canonical work pages

  1. [1]

    Arnav Arora, Lucie-Aim \'e e Kaffee, and Isabelle Augenstein. 2022. Probing pre-trained language models for cross-cultural differences in values. arXiv preprint arXiv:2203.13722

    arXiv 2022

  2. [2]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng X...

    arXiv 2023

  3. [3]

    Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell

    Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. https://doi.org/10.1145/3442188.3445922 On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, FAccT '21', pages 610--623, New York, NY, USA. Association for Comput...

    work page arXiv 2021

  4. [4]

    Zou, Venkatesh Saligrama, and Adam Kalai

    Tolga Bolukbasi, Kai - Wei Chang, James Y. Zou, Venkatesh Saligrama, and Adam Kalai. 2016. http://arxiv.org/abs/1607.06520 Man is to computer programmer as woman is to homemaker? debiasing word embeddings . CoRR, abs/1607.06520

    arXiv 2016

  5. [5]

    Christian Haerpfer, Ronald Inglehart, Alejandro Moreno, Christian Welzel, Kseniya Kizilova, Jaime Diez-Medrano, Marta Lagos, Pippa Norris, Eduard Ponarin, and Bjorn Puranen. 2021. World values survey: Round six - country-pooled datafile version. Available at: http://www.worldvaluessurvey.org/WVSDocumentationWV6.jsp

    work page 2021

  6. [6]

    Katharina Hämmerl, Björn Deiseroth, Patrick Schramowski, Jindřich Libovický, Alexander Fraser, and Kristian Kersting. 2022. http://arxiv.org/abs/2203.09904 Do multilingual language models capture differing moral norms?

    arXiv 2022

  7. [7]

    Masahiro Kaneko and Danushka Bollegala. 2021. http://arxiv.org/abs/2104.07496 Unmasking the mask - evaluating social biases in masked language models . CoRR, abs/2104.07496

    arXiv 2021

  8. [8]

    Teven Le Scao, Angela Fan, Christopher Akiki, Ellie Pavlick, Suzana Ili \'c , Daniel Hesslow, Roman Castagn \'e , Alexandra Sasha Luccioni, Fran c ois Yvon, Matthias Gall \'e , et al. 2023. Bloom: A 176b-parameter open-access multilingual language model

    2023

Show all 20 references
  1. [9]

    Moin Nadeem, Anna Bethke, and Siva Reddy. 2020. http://arxiv.org/abs/2004.09456 Stereoset: Measuring stereotypical bias in pretrained language models . CoRR, abs/2004.09456

    2020 arXiv

  2. [10]

    Moin Nadeem, Anna Bethke, and Siva Reddy. 2021. https://doi.org/10.18653/v1/2021.acl-long.416 S tereo S et: Measuring stereotypical bias in pretrained language models . In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th Inte...

    2021 doi

  3. [11]

    Nedjma Ousidhoum, Xinran Zhao, Tianqing Fang, Yangqiu Song, and Dit-Yan Yeung. 2021. https://doi.org/10.18653/v1/2021.acl-long.329 Probing toxic content in large pre-trained language models . In Proceedings of the 59th Annual Meeting of the Association for Computational Lingui...

    2021 doi

  4. [12]

    Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. https://api.semanticscholar.org/CorpusID:160025533 Language models are unsupervised multitask learners

    2019

  5. [13]

    Aida Ramezani and Yang Xu. 2023. Knowledge of cultural moral norms in large language models. arXiv preprint arXiv:2306.01857

    2023 arXiv

  6. [14]

    Nguyen, and Katrin Kirchhoff

    Julian Salazar, Davis Liang, Toan Q. Nguyen, and Katrin Kirchhoff. 2019. http://arxiv.org/abs/1910.14659 Pseudolikelihood reranking with masked language models . CoRR, abs/1910.14659

    2019 arXiv

  7. [15]

    Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A. Smith. 2019. https://doi.org/10.18653/v1/P19-1163 The risk of racial bias in hate speech detection . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1668--1678, ...

    2019 doi

  8. [16]

    Karolina Stanczak and Isabelle Augenstein. 2021. http://arxiv.org/abs/2112.14168 A survey on gender bias in natural language processing . CoRR, abs/2112.14168

    2021 arXiv

  9. [17]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. http://arxiv.org/abs/1706.03762 Attention is all you need . CoRR, abs/1706.03762

    2017 arXiv

  10. [18]

    Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. 2022. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068

    2022 arXiv

  11. [19]

    URL: " 'urlintro :=

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

  12. [20]

    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 12, 2026 · model on record in the stance chip above.