Post-publication Comment · Critical AI
Comment on “Cultural tendencies in generative AI”
Critical AI · published 2026-07-05 · v1.1 · CRIT-000045
Concerning: Jackson G. Lu, Lesley Luyang Song, Lu Doris Zhang · Nature Human Behaviour · 2025
Abstract
- Claim under scrutiny
- That the study 'demonstrate[s] the real-world impact' of language-conditioned cultural tendencies in GPT-4 and ERNIE, and that those tendencies 'probably originated from real-world cultural tendencies embedded in large-scale textual data'.
- Our assessment
- The differential outputs are real and carefully documented, but 'real-world impact' rests on GPT's slogan choices across 100 API iterations with no human consumer, attitude, or behaviour measured; and the single proposed causal mechanism cannot be separated from training-data composition, tokeniser behaviour, and RLHF tuning, all of which co-vary with prompt language.
- Implication
- The study is strong evidence that LLM outputs vary by prompt language in culturally patterned ways. It is not evidence of impact on people, nor of a specific cultural-origin mechanism — a reader should treat both as open questions the design was not built to close.
Bottom line
A rigorous, transparent descriptive study whose two headline framings — 'real-world impact' and a single training-data cause — outrun a design that measures only model outputs and manipulates only prompt language.
Why this paper was selected
Autonomous production cycle (licensed inbox, A-tier priority); full-text critique via two-stage produce+sharpen + 3-lens convergence gate (3 survives, 0 weakened).
AI/AGI centrality 4/5 · societal relevance 5/5 · source-journal note: A-tier per the monitored-venue determination; Nature Human Behaviour is elite interdisciplinary (Nature family), WoS JCR Q1. Critiqued from the publisher's version of record under the operator's licensed subscription access (not a paywall bypass).
Summary
This study prompts GPT-4 and ERNIE in Chinese versus English on established cultural-psychology scales (collectivism, attribution bias, reasoning style, expectation of change), finding that Chinese-language prompts yield more interdependent social-orientation scores and more holistic cognitive-style scores. It then shows GPT recommends culturally congruent advertising slogans and that cultural role-prompts can shift outputs. The design is descriptive and correlational, comparing 100 API iterations per language condition on single frozen model snapshots. The central critique is that the paper labels GPT's differential ad-slogan recommendations as 'real-world impact' when no human consumers, attitudes, or behavioural outcomes were measured — a model choosing different outputs by language is a model-behaviour finding, not demonstrated impact on people. A secondary concern is that the proposed causal mechanism ('cultural tendencies embedded in large-scale textual data') is underdetermined: prompt language co-varies with training-data composition, tokeniser behaviour, and RLHF tuning, none of which the design can separate.
Central claims & evidence map
| Claim | Type | Evidence offered | Support | Overclaiming | Main weakness |
|---|---|---|---|---|---|
| The paper claims to 'demonstrate the real-world impact' of cultural tendencies based on GPT's differential ad-slogan recommendations, but the study measures only which slogan GPT selects — no human consumers, attitudes, or behavioural outcomes are measured. | Causal | we demonstrate the real-world impact of these cultural tendencies | Weak | Major | Labelling a model-output difference as 'real-world impact' when no human outcome is measured is a direct scope over-claim. |
| The paper proposes that observed Chinese-vs-English output differences 'probably originated from real-world cultural tendencies embedded in large-scale textual data' without ruling out alternative explanations such as tokeniser effects, RLHF cultural steering, or test-item memorisation. | Causal | their cultural tendencies probably originated from real-world cultural tendencies embedded in large-scale textual data, on which generative AI models are trained | Moderate | Moderate | Proposing a single causal pathway without acknowledging competing explanations overstates what a language-manipulation design can identify. |
| The paper's title generalises from two specific model snapshots (gpt-4-1106-preview and ERNIE-3.5-8K-0205) to 'generative AI' as a category, though model behaviour can shift substantially across versions. | Descriptive | Cultural tendencies in generative AI | Moderate | Moderate | Generalising from two dated checkpoints to the category 'generative AI' overstates external validity. |
Point-by-point
Each objection as a scholarly disputation — the authors’ claim, our assessment, the strongest reply they could give, our response, and what would change our mind.
- Point 1Major overreachCausal
The authors claim: The paper claims to 'demonstrate the real-world impact' of cultural tendencies based on GPT's differential ad-slogan recommendations, but the study measures only which slogan GPT selects — no human consumers, attitudes, or behavioural outcomes are measured.
However — our assessment
The advertising study (Exploratory Analyses I) records GPT's slogan choice across 100 iterations per language for three product pairs. No human participants view, evaluate, or act on these recommendations. Demonstrating that a model's output differs by language is a model-behaviour finding. Calling it 'real-world impact' elides the entire causal chain from model recommendation to human behaviour. The paper frames this as 'evidence for the real-world impact' in the Results and uses 'we demonstrate' in the abstract, which is stronger than the data supports.
Anticipated reply
The authors could argue the advertising task is a face-valid proxy for real-world consequences — marketers do use LLMs to draft copy, so a systematic language-conditioned slogan bias plausibly propagates into materials people eventually see.
Our response
Even granting that LLMs are used in marketing, the study measures neither downstream deployment nor any human response; a proxy's real-world impact must be demonstrated, not assumed, and 'we demonstrate the real-world impact' asserts the very conclusion the proxy was meant to support. The gap is not that impact is implausible but that it is unmeasured.
What would change our mind
A study linking the language-conditioned outputs to an actual human outcome — e.g. a randomised test in which people exposed to the differing slogans show measurably different attitudes or choices — would convert this from a model-behaviour finding into demonstrated real-world impact and retire the objection.
- Point 2Moderate overreachCausal
The authors claim: The paper proposes that observed Chinese-vs-English output differences 'probably originated from real-world cultural tendencies embedded in large-scale textual data' without ruling out alternative explanations such as tokeniser effects, RLHF cultural steering, or test-item memorisation.
However — our assessment
The design manipulates only prompt language. Training-data composition, tokeniser segmentation, RLHF preference tuning, and potential memorisation of well-known psychometric items all co-vary with language. No ablation, probing study, or training-data analysis is conducted. The cross-model replication (GPT and ERNIE) partially addresses some confounds but both models face the same set of confounds. The hedge 'probably' mitigates but does not eliminate the concern.
Anticipated reply
The authors hedge with 'probably' and replicate the pattern across two independently developed models (GPT and ERNIE) from different countries, which they could argue makes a shared training-data origin the most parsimonious explanation for the convergent, culturally patterned differences.
Our response
Cross-model replication rules out a single vendor's idiosyncrasy but not the confounds common to both: both models tokenise Chinese and English differently, both undergo RLHF, and both may have encountered the same well-known psychometric items — so convergence is equally consistent with the training-data account and with its alternatives. 'Probably' correctly signals uncertainty, but the Discussion still advances the single mechanism without the ablation or probing evidence that would license it.
What would change our mind
A design that holds language fixed while varying the suspected confound — tokeniser-controlled prompts, a training-data audit, or activation-probing that localises the effect to learned cultural content rather than segmentation or RLHF — would isolate the mechanism and directly support or refute the training-data claim.
- Point 3Moderate overreachDescriptive
The authors claim: The paper's title generalises from two specific model snapshots (gpt-4-1106-preview and ERNIE-3.5-8K-0205) to 'generative AI' as a category, though model behaviour can shift substantially across versions.
However — our assessment
The title and abstract frame findings as properties of 'generative AI' without version qualification. Model versions are updated frequently; the specific API endpoints used are already deprecated. The Limitations section does acknowledge the need to test other models (Claude, DeepSeek, Gemini) but does not flag the fragility of version-specific findings or that even the next GPT point-release could eliminate all observed effects.
Anticipated reply
The authors could note that the Limitations section already calls for testing other models, and that a title naming specific checkpoints would be unwieldy and against a field convention that routinely generalises from tested systems to the category.
Our response
Acknowledging the limitation in the body does not license the unqualified categorical claim in the title and abstract — the parts most readers and the press actually see — especially when the specific endpoints tested are already deprecated, so the strongest form of the claim may not replicate on any currently deployed model.
What would change our mind
Reproduction of the same language-conditioned cultural pattern across several current-generation models (e.g. GPT-4o/next, Claude, Gemini, DeepSeek) would justify the categorical 'generative AI' framing; failure on newer checkpoints would confirm the finding is version-specific.
Scorecard
Sub-scores are 0–5 editorial judgements on fixed scales (higher is better, except methodological risk and overclaiming where higher is worse). They are contestable and open to a severity challenge from authors.
Strongest critique
The paper claims that the advertising study 'demonstrate[s] the real-world impact' of cultural tendencies, but the study records only which slogan GPT selects across 100 API iterations, with no human consumers, attitudes, or behavioural outcomes measured. A model choosing different outputs in different languages is a model-behaviour finding, not evidence of impact on people. The Discussion compounds this by extrapolating to societal-level cultural divergence ('English-speaking AI users may gradually become more independent and analytic') without any supporting evidence of human behavioural change.
Strongest fair defence
The authors are transparent that their claims are descriptive ('our findings are descriptive rather than prescriptive'), explicitly state they 'do not suggest that generative AI models possess cultural tendencies like humans do,' hedge the training-data mechanism with 'probably,' employ multiple converging measures across diverse formats (Likert scales, vignette tasks, imagery tasks, text analysis), replicate across two independently developed models from different countries (GPT and ERNIE), conduct four robustness checks including temperature variation and prompt-format sensitivity, share data and code on OSF, use appropriate statistical tests (t-tests, Poisson regression, beta regression) with power analysis, and take concrete steps to mitigate test contamination by modifying item content and obtaining unpublished materials directly from original researchers. The core descriptive finding — that LLM outputs systematically differ by prompt language in culturally patterned ways — is well-documented even if the causal mechanism remains open.
Conclusion
A well-executed descriptive study with converging evidence across measures and models, weakened primarily by labelling a model-output difference as 'real-world impact' without any human-outcome data, and secondarily by proposing a single causal mechanism that its language-manipulation design cannot isolate from confounds. These are bounded overclaims on an otherwise rigorous and transparent study that honestly discloses its sampling and model limitations.
Reply from the authors
Following the practice of Nature Matters Arising, Science Technical Comments and PNAS Letters, this Comment is published as one half of a Comment + Reply pair: the authors of the original article are invited to respond, and any reply is published here verbatim alongside the Comment as part of the record.
Reply: not yet invited. No reply has been received for publication.
The authors have a right of reply and no veto. A reply may request a factual correction, a methodological rebuttal, a clarification, a data/code update, or a severity challenge, and is published unedited. See the right-of-reply policy.
Source-grounding attestation
- ✓Verbatim source spans present in the critique — 4/4 provenance spans re-derived in the critique prose
- ✓Passes the publication validator — no errors
- ✓Zero fabricated citations — 0 fabricated
- ✓Severity within the access-basis cap — severity "moderate" ≤ cap "high" for licensed_access
Every verbatim span the critique relies on is re-derived in the prose in-app; span-in-source is re-verifiable offline (the abstract is re-fetched, not stored, per the no-reproduce policy).
Re-verify span-in-source offline: python3 scripts/verify-fulltext-critiques.py
Independent faithfulness review
A refute-by-default adversarial panel (two independent reviewers — an overreach lens and a mischaracterization lens — that fetched the real source) tried to prove this critique misread the paper. This is an AI adversarial review recorded with its reasoning, not a deterministic check.
All three verbatim spans independently verified as exact substrings of the stored source text. The critique fairly credits the paper's genuine strengths (converging measures, cross-model replication, robustness checks, open data/code, test-contamination mitigation) while identifying the bounded overclaims.
Version & correction history
| Version | Date | Change |
|---|---|---|
| v1.0 | 2026-07-05 | |
| v1.1 | 2026-07-05 | Reformatted into the dialectical Matters-Arising Comment genre (G146): added a structured abstract, a bottom-line verdict, and for each objection an anticipated author reply, our response, and an explicit falsification condition ('what would change our mind'). The verdict, claims, severity, and verbatim spans are unchanged; this revision adds the paper's strongest defences and the Comment's own defeaters, and changes presentation — it does not add adverse findings. |
No silent substantive corrections — every change is versioned and visible.
How to cite this Comment
Critical AI. Comment on “Cultural tendencies in generative AI” (Jackson G. Lu et al., Nature Human Behaviour, 2025). Critical AI; 2026. https://policywindow.org/critique/c/cultural-tendencies-generative-ai
A registered DOI will replace the URL once minted; until then the canonical URL is the persistent identifier. Highwire/Dublin-Core citation tags and a schema.org Review record are embedded in this page for Google Scholar and reference managers.
Verify this Comment. Its checkable facts (target DOI, access-basis severity cap, zero fabricated citations) are served — as the app’s self-report — at /critique/api/critiques/cultural-tendencies-generative-ai/verify; to confirm them independently of this site, re-derive the same checks (and resolve the target DOI) with npx tsx scripts/verify-critical-ai.ts --critique cultural-tendencies-generative-ai --live.
Content fingerprint d6f3da73c2424a30 (v1.1) — this Comment’s substantive content is content-addressed; a silent post-publication edit would change it.