PaperlayerView exact revision on arXiv

Source-linked AI summary

LLMs Mirror Country-Specific Gender Patterns If Asked, but Skew Male When Generating Media in Local Languages

Sharif Kazemi, Tanya Popli, Neil K. R. Sehgal, Sunny Rai, Niyati Malhotra, Victor Orozco-Olvera, Ana María Muñoz Boudet, Samuel P. Fraiberger, Sharath Chandra Guntuku, Manuel Tonneau

arXiv:2609.06545v1cs.CL

TL;DR

Selection-based evaluations and scarce local human baselines leave open how LLMs express gender associations in generated media across countries and languages. The paper surveys role associations in four countries and compares eight models’ direct answers with generated media, finding a male shift concentrated in major local-language cells.

  • Problem

    Selection-based evaluations do not reliably capture long-form generation, and human baselines for local gender associations are scarce outside Western contexts.

    Source: §1 · p. 2

  • Method

    The study surveys gender associations for 22 occupational and domestic roles among respondents in four countries, then evaluates eight LLMs through direct questioning and media generation.

    Source: §1 · p. 2, §3.1 · p. 3, §4.2 · p. 5

  • Results

    Models broadly track surveyed associations when questioned directly but shift male in media generation, especially in major local-language cells.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Evaluations of LLM gender bias for global deployment should include generation-format testing, local-language prompting, and locally collected human baselines.

    Source: §7 · p. 8

    The mitigation experiments lack a non-gender-related placebo control, so they cannot isolate gender-instruction effects from shifts caused by adding any instruction.

    Source: §8 · p. 9

Abstract

from arXiv · show

Large language models (LLMs) are increasingly used to generate media, but whether their content perpetuates gender stereotypes is unknown: standard benchmarks rely on selection-based formats rather than long-form generation, and surveyed baselines for local gender associations are scarce outside the West. We collect gender associations for 22 occupational and domestic roles from 695 respondents across the United States, India, Kenya, and Nigeria, and evaluate eight LLMs under two regimes: direct questioning and media generation. Models track the surveyed associations under direct questioning but skew substantially more male under media generation in major local-language cells, consistent with the male bias documented in human-produced media. Outside the US, the shift is much smaller and non-significant under English prompting, so English-only or country-agnostic evaluation would miss this bias in the languages where these models are most deployed. Instruction prompting reduces the shift directionally, but trades off against alignment with the surveyed associations. Evaluating LLM gender bias for global deployment therefore requires generation-format testing, local-language prompting, and locally-collected human baselines.

1 Introduction

The paper addresses gaps in generation-based evaluation and local human baselines by surveying gender associations in four countries and comparing them with eight LLMs. Models broadly track surveyed associations when questioned directly but shift male in media generation, especially in major local-language cells; instruction-based mitigations are brittle.

1 Introduction

  • Selection-based evaluations and scarce non-Western human baselines leave unclear how LLMs behave in long-form generation across deployment contexts.Existing selection-based metrics do not reliably correlate with generation-based evaluations.Source: §1 · p. 2
  • The study collects gender associations for 22 occupational and domestic roles in India, Kenya, Nigeria, and the United States to provide a human comparator for eight LLMs.The survey is validated against external labour statistics where available.Source: §1 · p. 2
  • Models broadly track surveyed associations when questioned directly but shift male in generated media, with significant shifts in major local-language cells and smaller, non-significant shifts under English prompting outside the US.The shift is roughly half a Likert step and converges with the approximately 75% male share documented in film and television analyses.Source: §1 · p. 2
  • Instruction-based mitigation is directionally effective but brittle, with intuitive variants overshooting toward female.These variants trade reduced male shift for weaker alignment with surveyed associations.Source: §1 · p. 2

2 Related Work

Prior work identifies limitations in selection-style bias benchmarks and gaps in cross-cultural, language-sensitive evaluation. Research on media bias and LLM generation also leaves long-form entertainment media comparatively underexplored.

LLM bias evaluation

  • Selection-based bias benchmarks can under-predict long-form generation behavior, while explicit questionnaires may not reveal implicit associations surfaced by contextualized generation.Prior findings include mixed gender-occupation results across Western generation audits.Source: §2 · p. 2

Cross-cultural evaluation of LLMs

  • Cross-cultural audits and local baselines remain limited, and prompting language itself is rarely evaluated despite evidence that in-language elicitation can surface patterns invisible in English.The paper addresses these gaps with gender-association baselines across four countries and evaluations in five prompt languages.Source: §2 · p. 2

Media bias

  • Although studies document gender stereotypes in text and image generation, long-form entertainment media remains an underexplored deployment context.Prior work reports both media stereotypes and downstream effects on viewer aspirations.Source: §2 · p. 3

3 Data

The paper builds a four-country survey of perceived gender associations for occupational and domestic roles, then uses independent media analyses as a comparison baseline. The survey captures cross-country differences and is validated against available external benchmarks.

3.1 Cross-country survey on gender role association

  • The online survey gathered 695 adults across India, Kenya, Nigeria, and the United States, using self-selected English or local-language participation rather than a probability sample.The US sample was deliberately smaller; respondents were selected from the internet-connected population most likely to use LLMs in each country.Source: §3.1 · p. 3
  • The survey asks descriptive questions about 12 occupations and 10 domestic activities, scoring responses from −2 for male-associated to +2 for female-associated and excluding “Don’t know.”The same scoring is applied to LLM responses.Source: §3.1 · pp. 3–4
  • Respondents agree on strongly gendered roles, while India differs from the other countries on several service-sector roles.Elder care is consistently female-coded and engineering and handicraft work male-coded; hairdresser, secretary, and waiter/bartender associations diverge in India.Source: §3.1 · p. 4
  • The survey correlates with US occupational statistics at r = +0.95 and with Indian formal-sector statistics at r = +0.89, but its overall Indian occupational correlation is r = +0.42.Comparable statistics were unavailable for Kenya and Nigeria; domestic-task benchmarks correlate at r = +0.94 across eight matched pairs.Source: §3.1 · p. 4

3.2 Gender representation in media

  • Independent film and television analyses provide media baselines; generated implicit male shares converge with their documented male over-representation.The cited analyses report 80.5% male working characters in US film and television and an approximately 75% male share across film and television analyses.Source: §3.2 · p. 4

4 Experimental Design

The experiment compares explicit role judgments with implicit media generation while varying models, nationality framing, and prompt language. Both regimes use the same roles and human comparator, enabling an estimate of how elicitation format changes directional gender scores.

4 Experimental Design

  • The factorial design varies model, elicitation regime, nationality framing, and prompt language to compare selection-style evaluation with open-ended generation under a constant role set and human comparator.Its factors are intended to distinguish each contribution to the observed contrast.Source: §4 · p. 5

4.1 Models

  • Eight models include one frontier closed model, four generic open-weight models, and three country-targeted language-localized fine-tunes.The localized models target India, Kenya, and Nigeria.Source: §4.1 · p. 5

4.2 Elicitation regimes

  • Explicit elicitation asks the model to answer the survey question in free text, which a GPT-4o-mini judge maps onto the survey’s directional scale.The judge’s headline result is robust to two open-weight judges and a regex classifier.Source: §4.2 · p. 5
  • Implicit elicitation asks the model to generate media with an unspecified protagonist, then measures that protagonist’s gender from the response.Prompts cover five media types, four nationality framings, 22 roles, and five languages; each combination has five paraphrases and two seeded runs.Source: §4.2 · p. 5
  • The primary analysis uses OLS on response-level directional scores, with fixed effects for model, task, and nationality framing and standard errors clustered by task.The implicit indicator’s coefficient estimates the score change from explicit to implicit elicitation while holding those factors fixed.Source: §4.2 · p. 6
  • The pre-specified secondary test examines whether the elicitation shift differs by prompt language within country; model-, cell-, and task-level breakdowns are descriptive.The analysis reports Benjamini–Hochberg false-discovery-rate q-values alongside unadjusted p-values for each family.Source: §4.2 · p. 6

5 Results

Across the eight models, media generation shifted gender representation male relative to direct questioning. The shift was amplified by local-language prompting, while instruction variants did not reliably mitigate it without compromising alignment with surveyed associations.

5.1 LLM gender associations

  • Implicit elicitation lowered alignment by roughly 0.23 on average (β = −0.72 on [−2, +2], 95% CI [−1.10, −0.34], p < .001), from explicit mean Pearson r = +0.81.Collapsing explicit responses to the implicit scale preserved the direction and roughly 75% of the estimated magnitude.Source: §5.1 · p. 6
  • Six of eight models shifted significantly male from explicit to implicit elicitation, and implicit outputs exceeded survey male shares on 19 of 22 roles.Implicit elicitation averaged 66% male, compared with 44% for explicit elicitation and a 75% male film/TV reference; the comparison is directional.Source: §5.1 · p. 6, Figure 3 · p. 7

Country-language interaction on male shift

  • Local-language prompting produced four-to-six-times larger shifts in India and Nigeria and roughly doubled Kenya’s, while Hausa showed no effect; English shifts were small and non-significant.English-only evaluation would understate the deployment-relevant shift in three of the four countries.Source: Country-language interaction on male… · pp. 6–7, Figure 4 · p. 7

5.2 Mitigations to male shift

  • Balanced and debias instructions overshot toward female, lowering GPT-5.1 alignment from r = +0.45 to −0.22 and −0.09; realism brought no reliable gain.The variants were tested across all eight models, with sampling held constant and four nationality framings cycled.Source: §5.2 · p. 7, Figure 5 · p. 7

6 Discussion

The discussion frames the male shift as context-specific and consistent with media-derived training patterns. It also argues that prompt mitigations trade off against surveyed alignment and cannot resolve the normative target for deployment.

Implications for evaluation

  • The gap aligns with claims that selection-based benchmarks under-predict generation behavior (Lum et al., 2025; Goldfarb-Tarrant et al., 2021; Cheng et al., 2023), motivating generation-format evaluation for media use.The authors emphasize entertainment-media cases where a protagonist’s gender is inferred rather than specified.Source: §6 · p. 8

Context-specific male shift

  • Implicit elicitation exceeded human-survey male shares on 19 of 22 roles and approached the ≈75% male share in film and television, a pattern the authors frame as representational.They suggest models reflect media-derived training corpora and may extend at scale harms associated with human-authored media.Source: §6 · p. 8
  • The authors attribute differences from context-agnostic storytelling to role, nationality, and language anchoring, as well as pronoun- versus name-based gender extraction.They also connect UlizaLlama3’s opposite shift to its fine-tuning conversations dominated by pregnant women and new mothers (Mwongela et al., 2024; Jacaranda Health, 2024).Source: §6 · p. 8

Brittle mitigation and the alignment-target question

  • Survey-tuning reproduces patterns respondents currently exhibit, while balanced targets encode developers’ values; choosing between these normative frames belongs to deployers, policymakers, and affected communities.The authors describe their own role as documenting current behavior rather than selecting the target.Source: §6 · p. 8

7 Conclusion

Across eight models, explicit questioning tracks surveyed gender associations, while implicit media generation lowers alignment and shifts outputs male, especially in local-language settings. The shift is smaller and non-significant under English prompting outside the US, and instruction prompting is directionally effective but brittle.

7 Conclusion

  • Across four countries and five prompt languages, eight models track the 695-respondent survey under explicit elicitation but shift male under implicit media generation, especially in local languages.The shift is non-significant under English prompting outside the US; simple instruction prompting is directionally effective but brittle.Source: §7 · p. 8
  • Survey-style benchmarks understate deployment-time representational behavior, particularly in non-Western settings where local human baselines are scarce.The conclusion calls for evaluation in the format models are actually used in and a deliberate choice about the alignment target.Source: §7 · p. 8

8 Limitations

The study's limitations include reliance on an LLM judge, a coarse implicit-outcome scale, and a narrow evaluation of instruction prompting. Its survey and language coverage also represent selected online populations and a subset of commonly spoken languages.

8 Limitations

  • GPT-4o-mini serves as the judge for free-text scoring and gender extraction, though the extraction classifier was validated against human annotation.The reported extraction macro-F1 is 0.97 for single-protagonist cases and 0.75 end-to-end with narrative gating.Source: §8 · p. 9
  • The ternary media-generation scale loses within-category intensity and combines collectivist narratives, format mismatches, and safety refusals as unknown.Collapsing explicit responses to the same three-point support preserves the headline coefficient's direction and most of its magnitude, though significance is marginal under the strictest inference.Source: §8 · p. 9
  • The mitigation study tests four instruction variants on eight models without a non-gender instruction control, so instruction-content effects cannot be separated from adding an instruction.The authors treat these findings as descriptive evidence of unreliable simple prompt interventions, not a comprehensive mitigation evaluation.Source: §8 · p. 9
  • Facebook-Messenger recruitment skews younger, more urban, and more educated within countries, so the survey reflects this stratum rather than each country's population uniformly.The authors note that this internet-connected group may be especially likely to use LLMs for media generation.Source: §8 · p. 9
  • The evaluation includes at least one major local language per country, but does not cover the breadth of commonly spoken languages or many linguistic subgroups.This bounds the linguistic coverage of the study's findings.Source: §8 · p. 9
  • The prompts differ in structure across elicitation regimes, despite averaging five implicit paraphrases per cell and five media types per role.The explicit prompt asks directly about a role, whereas implicit prompts request generated media featuring a protagonist performing it.Source: §8 · p. 9

A.1 Per-language yields and demographic profile

  • Respondents self-selected survey language without quotas, and local-language samples reflect people who use those languages online.Local-language yields ranged from 41 Yoruba respondents to 120 Swahili respondents; English recruited more respondents in every country.Source: §A.1 · p. 13

A.2 Role selection

  • The 22 roles were selected with expert consultation, balancing paid and unpaid work and drawing domestic activities from time-use survey instruments.The occupations span formal and informal sectors; the formal-statistics validation is reported separately in Figure 6.Source: §A.2 · p. 13, §A.3 · p. 13

A.4 Survey validation against household-task benchmarks

The appendix reports validation of the human survey against household-task benchmarks, demographic robustness checks, prompt and model details, and validation of the scoring pipeline. The household-task survey associations closely match external benchmarks, and alternative scoring classifiers largely agree on direction.

A.5 Demographic robustness across respondent strata

  • For the three larger country samples, female-only and male-only respondent alignment intervals overlap substantially across all country-model combinations.These comparisons were made against the all-respondents baseline.Source: §A.5 · p. 13

B.1 Media-content baseline sources

  • The media-content baselines document male-skewed occupational representation in US television and films, cross-country feature films, Bollywood scripts, and Nollywood films.Reported examples include 80.5% male working characters in US prime-time television and family films, and only 18% of female Nollywood characters in remunerated employment.Source: §B.1 · pp. 13–14

B.3 Prompt examples per language

  • The prompt examples hold the role to Medical Doctor across five languages, with implicit examples drawn from one of five paraphrase variants.Local-language translations were made by native speakers and spot-checked for fidelity.Source: §B.3 · pp. 14–15

A.4 Survey validation against household-task benchmarks

  • Across eight matched household-task pairs in four countries, survey associations correlate with benchmarks at r = +0.94, with MAE = 0.34 on the ±2 scale and 100% sign agreement.The benchmarks combine research sources including household earnings, unpaid eldercare, household-task polling, and demographic and health surveys.Source: §A.4 · p. 13, Figure 7 · p. 15

B.2 Models evaluated

  • The models table distinguishes localisation by curated supervised fine-tuning, authentic in-language conversational data, or machine-translated supervision.These categories are labeled curated SFT, community, and MT-SFT.Source: Table 2 · p. 16

B.3.1 Explicit elicitation prompt template

  • Explicit prompts use a common English framing and answer guide, insert the role question in the target language, and map free-text replies to a [−2, +2] score.GPT-4o-mini performs the scoring.Source: §B.3.1 · p. 16

B.3.2 Implicit elicitation prompts

  • Implicit prompts are fully translated and ask models to generate media featuring an unspecified-gender protagonist performing the role.Five variants alter wording and syntax while preserving the role and country slots; the variant pools were generated once and not regenerated per evaluated model.Source: §B.3.2 · pp. 16–17

C.1.1 Per-classifier agreement with GPT-4o-mini

  • Alternative classifiers agree with GPT-4o-mini on direction in 96.8–98.7% of compared responses, with disagreements concentrated at the boundary between ±1 and ±2 scores.The pipeline uses GPT-4o-mini for both explicit scoring and implicit gender extraction, with each stage validated separately.Source: §C.1.1 · p. 17, C Pipeline validation · p. 17

C.1.2 Per-model cross-classifier agreement

The appendix reports a male-direction implicit shift across models, with substantial variation by country-language cell. Checks using alternative classifiers, human annotators, and different handling of unclassifiable responses preserve the direction of the headline result.

C.1.2 Per-model cross-classifier agreement

  • Across evaluated models, alternative judges retain sign-plus-zero agreement above 87%; localized fine-tunes are hardest to classify, and N-ATLaS under Qwen 3 is lowest at 87.1%.Agreement across language sheets exceeds 93% for both open-weight judges, with lower exact agreement on local-language sheets.Source: §C.1.2 · p. 17, Table 4 · p. 18, Table 5 · p. 19

C.1.4 Downstream impact of substituting an alternative classifier

  • Replacing GPT-4o-mini's explicit-response scores with alternative classifiers never flips any model's implicit-shift coefficient; the regex classifier yields the smallest magnitude.The point estimate shrinks modestly under conservative classifiers, and significance is retained except under the strictest threshold.Source: §C.1.4 · p. 18, Table 6 · p. 19

C.2 Implicit gender extraction: human-annotator validation

  • On 71 responses where both human annotators assigned a specific gender, GPT-4o-mini matches their consensus on 97.2% (69/71).Human annotators agreed on 83.1% overall; disagreement primarily concerned whether to assign a gender or mark unknown.Source: §C.2 · p. 18

C.3 Collectivist narratives: handling and sensitivity

  • The male shift remains significant when non-individualistic implicit responses are excluded, including under the strictest filter requiring a successfully extracted individual protagonist.Coding unclassifiable responses as zero conflates collectivist narratives, format mismatches, refusals, and extraction errors.Source: §C.3 · p. 19, Table 7 · p. 19

D.2.1 Per (country × language) cell coefficient

  • Overall, switching to implicit elicitation yields β = −0.719 across n = 37,707 observations; six models drift male, while UlizaLlama3 drifts female with a narrowly non-significant effect.The figure reports model-level marginal effects with 95% confidence intervals.Source: Figure 9 · p. 20

D.2.2 Within-country English-vs-local interaction

  • India × Hindi and Nigeria × Yoruba show the largest significant male shifts; Kenya × Swahili also amplifies over English (∆β = −0.40, p = .001), while Hausa shows no amplification (∆β = −0.06, p = .67).The largest Global South English-with-country-framing shift is −0.37 for Kenya, about a quarter of the largest local-language shift, Hindi at −1.36.Source: §D.2.1 · p. 20, §D.2.2 · p. 21

D.3.1 Variance components

  • The implicit shift in mean score is 0.64, about six times between-model SD and larger than between-country-framing or between-language SD; country × language SD is 0.44.Across the eight country-language cells, three local-language cells drift markedly male, while the four English cells cluster near neutral and Hausa does not amplify.Source: §D.3.1 · p. 21, Table 11 · p. 22, §D.3.2 · p. 22

D.3.3 Within-paraphrase variance

  • Implicit paraphrase variation has mean within-cell SD 0.93 Likert units across 4,510 cells, while pooled task shifts are male-directed on 15 of 22 roles.The largest male shifts occur on female-coded service and care roles; seven tasks shift slightly female where explicit responses already concentrate near the male ceiling.Source: §D.3.3 · p. 22, §D.3.4 · p. 22

D.3.5 Per (model × task) cell extremes

Across 176 model-task cells, shifts more often moved toward male, with Cook and Playing with children concentrated among the strongest male shifts. A task-cluster wild bootstrap raises the headline p-value but leaves it below conventional thresholds.

D.3.5 Per (model × task) cell extremes

  • 117 of 176 cells (66.5%) shifted toward male, versus 59 (33.5%) toward female; the median shift was −0.71 Likert units.The mean was −0.60, and shifts ranged from −2.56 to +1.55.Source: §D.3.5 · p. 22
  • Cook and Playing with children concentrated among the most male-shifted cells; LLaMA 3.1 8B on Cook moved from a near-pure female baseline to a clearly male narrative.Reverse shifts included seven tasks whose explicit baseline was at or below neutral, leaving movement toward neutrality as the available headroom.Source: §D.3.5 · pp. 22–23, Table 13 · p. 23

E.1 Small-sample inference: wild cluster bootstrap by task

  • The task-cluster wild-bootstrap p-value was 0.0018 versus 0.0002 asymptotically, but both remained below conventional thresholds and the point estimate was unchanged.The bootstrap used 4,999 Rademacher draws to address small-cluster inference with 22 task clusters.Source: §E.1 · p. 23

E.2 Scale-support robustness: collapsing the explicit scale to three points

Collapsing explicit scores to the implicit regime’s three-point support retained roughly three-quarters of the estimated shift, though the estimate became marginal under the wild bootstrap. The paper therefore treats the five-point specification as primary and the collapsed result as a conservative lower bound.

E.2 Scale-support robustness: collapsing the explicit scale to three points

  • The scale-collapsed estimate retained roughly 75% of the original magnitude—about half a Likert step—but was only marginally distinguishable from zero (wild-bootstrap p = .053).Both specifications used the same free-text-only sample (n = 37,707); the five-point specification remains primary.Source: §E.2 · p. 23, Table 16 · p. 24
  • Collapsing scores maps explicit responses of −2/−1 to −2, 0 to 0, and +1/+2 to +2, matching the implicit coding’s three-point support.This re-coded 53% of explicit responses that were at ±1; the implicit side was unchanged.Source: §E.2 · p. 23
  • The recoding narrows the contrast because explicit responses at ±1 skew more male than female, moving the explicit mean from +0.24 to +0.08.The paper reads the collapsed specification as a conservative lower bound because the recoding discards intensity information and widens uncertainty.Source: §E.2 · p. 23
Loading 2609.06545v1…