Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat
Sign Up
Cited by 29Ethics and Social Impacts of AIPsychology of Moral and Emotional JudgmentCultural Differences and Values

The DOI matches its Crossref record; 88 of the 88 references it cites were found in a registry or an index.

Moral stereotyping in large language models

Open access
Zewail, Aliah, Figueroa, Alexandra, Graham, Jesse, Atari, Mohammad·2026·Proceedings of the National Academy of Sciences
DOI 10.1073/pnas.2519941123PMC PMC12974530
Open this paper in CoChat — free account

Abstract

Can Large Language Models (LLMs) accurately estimate various societies’ moral values? Here, we query the perceptions of LLMs regarding the moral norms of the “average” person from 48 nations and compare them to a large-scale ( n = 90 , 802 ) survey of six moral values (Care, Equality, Proportionality, Loyalty, Authority, and Purity) from those populations. Our findings indicate that LLMs poorly capture the moral diversity around the globe, systematically overestimating some moral values (particularly Care) and underestimating others (especially Purity). Notably, examining various versions of Generative Pre-trained Transformer (GPT) shows that these LLMs may overestimate the overall moral concerns of some Western countries (e.g., the United States, Canada, and Australia) while underestimating those of non-Western countries (e.g., Nigeria, Morocco, and Indonesia). Our work demonstrates that LLMs are inaccurate generators of cross-cultural estimations in the moral domain; in other words, they stereotype the moral values of non-Western populations in predictable ways. Our results highlight the ethical and epistemic risks of relying on LLMs to estimate the endorsement of moral values around the globe.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1073/pnas.2519941123

Cites

88 of 88 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1. The difference between the moral value estimate (produced by GPT-3.5, GPT-4, and GPT-4o) and the ground truth (human sample). Red coloration represents an overestimation of moral values, and blue coloration represents an underestimation of moral values.
Fig. 1. The difference between the moral value estimate (produced by GPT-3.5, GPT-4, and GPT-4o) and the ground truth (human sample). Red coloration represents an overestimation of moral values, and blue coloration represents an underestimation of moral values.CC BY-NC-ND · PubMed Central
Fig. 2. The map above illustrates, overall, how much GPT-4o’s moral endorsements deviate (i.e., distance score) from the actual participant scores for each country.
Fig. 2. The map above illustrates, overall, how much GPT-4o’s moral endorsements deviate (i.e., distance score) from the actual participant scores for each country.CC BY-NC-ND · PubMed Central
Fig. 3. The plots above visually break down the correlations between difference scores and cultural distance scores by the GPT model. Also, GPT generally tends to overestimate four values (Care, Proportionality, Loyalty, and Authority) while underestimating two values (Equality and Purity).
Fig. 3. The plots above visually break down the correlations between difference scores and cultural distance scores by the GPT model. Also, GPT generally tends to overestimate four values (Care, Proportionality, Loyalty, and Authority) while underestimating two values (Equality and Purity).CC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 82 references

Research works printed but not linked: 2

  1. L. Hewitt A. Ashokkumar I. Ghezae R. Willer Predicting results of social science experiments using large language models. DocSend [Preprint] (2024). https://docsend.com/view/ity6yf2dansesucf. Accessed 20 November 2025.
  2. 10.1037/a0021847 · DOI · Paper page

Cited by

Related

See all

Metadata

Authors
Zewail, A., Figueroa, A., Graham, J., & Atari, M.
Affiliations
  • University of Massachusetts Amherst (US)
  • University of California System (US)
  • University of Utah (US)
Corresponding author
Atari, Mohammad
ORCID
Venue
Proceedings of the National Academy of Sciences
Publisher
National Academy of Sciences
Published
2026-03-04
Volume
123
Issue
10
Pages
e2519941123-e2519941123
ISSN
  • 0027-8424 (ISSN-L)
  • 1091-6490
DOI
10.1073/pnas.2519941123
PMC
PMC12974530
OpenAlex
W7133518443
Semantic Scholar
cf62c036ba481db69ca3896f867f052ad652005d
Cited by
29
References
82
Citations by year
2026: 21
FWCI
29.83
Type
journal article
Language
en
Open access
Licence
Locations
  • Proceedings of the National Academy of Sciences (publishedVersion, cc-by-nc-nd)
  • PsyArXiv (OSF Preprints) (submittedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
Funders
National Science Foundation (2439846)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Psychology of Moral and Emotional Judgment
  • Social Sciences › Psychology › Social Psychology › Cultural Differences and Values
Keywords
large language models, moral values, moral foundations, GPT, cultural bias in AI, Western bias
MeSH terms
Large Language Models, Generative Artificial Intelligence, Cross-Cultural Comparison, Humans, Morals, Stereotyping

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.