The Male Abstraction Bias in Internet Culture and Large Language Models
Organizations: Stanford University, Stanford, CA, USA · Oxford Internet Institute, Oxford, UK
Abstract
Conceptual abstraction and gender are two equally central aspects of the human mind, yet they have traditionally been studied in isolation from one another, as though they were fundamentally separate. Here, we find gender and abstraction are systematically linked throughout the space of concepts spanning single (N = 40,000) and multi-word expressions (N = 67,000): across billions of words and millions of images on the internet --- and across numerous measurement techniques, including human and AI classifications --- more abstract concepts are more associated with men, a pattern we show replicates in languages around the world. This male abstraction bias not only holds across semantic domains, but it also maps onto grammatical structure: verbs and adverbs are more abstract and more male-coded, revealing a system-wide trace of stereotypes accentuating agentic qualities in men. Furthermore, we show this bias readily emerges in the free associations humans and AIs produce in controlled online tasks, indicating its psychological internalization and its algorithmic reproduction. These findings identify a latent structure in conceptual space that unifies a range of stereotypes: for example, framings of women as more embodied, emotional, and warm, versus men as more cognitive, rational, and competent; as well as erroneous views of men as better at mathematics, science, and engineering. More broadly, by analyzing an unprecedentedly large, multimodal sample of online data, we provide culture-wide evidence of gender as a core organizing principle shaping conceptual space far beyond the characterization of people and social roles.
Figures & tables
| Primary axis (5 pairs) | Extended axis (10 pairs) | |||||
|---|---|---|---|---|---|---|
| Specification | Estimate [95% CI] | Estimate [95% CI] | ||||
| No controls (Supplementary Fig. 1 ) | [ , ] | 23,758 | [ , ] | 23,758 | ||
| Excluding lexically gendered terms and seed words | [ , ] | 23,062 | [ , ] | 23,062 | ||
| Excluding the 5% most strongly gendered terms | [ , ] | 22,570 | [ , ] | 22,570 | ||
| Excluding the 10% most strongly gendered terms | [ , ] | 21,382 | [ , ] | 21,382 | ||
| , controlling for word frequency | [ , ] | 23,407 | [ , ] | 23,407 | ||
| Axis | Pairs | [95% CI] | Most feminine words | Most masculine words | |
| A. Axes built from several unambiguous kinship/role pairs | |||||
| Primary set (5 pairs) | 5 | [ , ] | babydoll , bruidsmeisje , lesbienne , brunette , topless | aanvoerder , voorman , voorganger , manschappen , bevelhebber | |
| Extended set (10 pairs) | 10 | [ , ] | babydoll , brunette , blondine , lingerie , tienermeisje | leenheer , voorman , aanvoerder , veldheer , hopman | |
| Primary set without man / vrouw | 4 | [ , ] | babydoll , brunette , topless , bruidsmeisje , pinup | aanvoerder , voorman , voorganger , gebroeders , meesterzet | |
| Primary set without jongen / meisje | 4 | [ , ] | babydoll , lingerie , seksbom , brunette , lesbienne | voorman , veldheer , kompaan , aanvoerder , opportunist | |
| Primary set without mannelijk / vrouwelijk | 4 | [ , ] | lesbienne , babydoll , topless , brunette , bruidsmeisje | aanvoerder , voorman , voorganger , veedrijver , oudgediende | |
| Primary axis (5 pairs) | Extended axis (10 pairs) | |||||
|---|---|---|---|---|---|---|
| Specification | Estimate [95% CI] | Estimate [95% CI] | ||||
| No controls (Supplementary Fig. 2 ) | [ , ] | 33,031 | [ , ] | 33,031 | ||
| Excluding lexically gendered terms and seed words | [ , ] | 32,263 | [ , ] | 32,263 | ||
| Excluding the 5% most strongly gendered terms | [ , ] | 31,379 | [ , ] | 31,379 | ||
| Excluding the 10% most strongly gendered terms | [ , ] | 29,727 | [ , ] | 29,727 | ||
| Rater s.d. of concreteness median (reliably rated words) | [ , ] | 16,609 | [ , ] | 16,609 | ||
| Axis | Pairs | [95% CI] | Most feminine words | Most masculine words | |
| A. Axes built from several kinship/role pairs | |||||
| Primary set (5 pairs) | 5 | [ , ] | sõbranna , õde , kleit , pulmakleit , rootslanna | vend , roolimees , raudmees , ülem , pealik | |
| Extended set (10 pairs) | 10 | [ , ] | sõbranna , tädi , rootslanna , kaaslanna , eksnaine | vend , onu , pojapoeg , vilemees , poeg | |
| Primary set without mees / naine | 4 | [ , ] | kleit , sõbranna , õde , meik , pulmakleit | vend , poeg , pojapoeg , pealik , isa | |
| Primary set without poiss / tüdruk | 4 | [ , ] | õde , ämmaemand , sõbranna , kleit , eksnaine | vend , raudmees , roolimees , mängumees , vilemees | |
| Primary set without mehelik / naiselik | 4 | [ , ] | õde , sõbranna , meditsiiniõde , kleit , ämmaemand | vend , roolimees , pealik , vilemees , tehnikadirektor | |
| Primary axis (5 pairs) | Extended axis (10 pairs) | |||||
|---|---|---|---|---|---|---|
| Specification | Estimate [95% CI] | Estimate [95% CI] | ||||
| No controls (Supplementary Fig. 3 ) | [ , ] | 23,961 | [ , ] | 23,961 | ||
| Excluding lexically gendered terms and seed words | [ , ] | 23,559 | [ , ] | 23,556 | ||
| Excluding the 5% most strongly gendered terms | [ , ] | 22,762 | [ , ] | 22,762 | ||
| Excluding the 10% most strongly gendered terms | [ , ] | 21,564 | [ , ] | 21,564 | ||
| Rater s.d. of concreteness median (reliably rated words) | [ , ] | 12,152 | [ , ] | 12,152 | ||
| Axis | Pairs | [95% CI] | Most feminine words | Most masculine words | |
| A. Axes built from several kinship/role pairs | |||||
| Primary set (5 pairs) | 5 | [ , ] | 少妇, 美女, 女人, 少女, 女性 | 运气, 长兄, 如何, 数值, 设法 | |
| Extended set (10 pairs) | 10 | [ , ] | 美女, 少妇, 女人, 女性, 少女 | 长兄, 兄长, 他们, 继位, 晋见 | |
| Primary set without 男人/ 女人 | 4 | [ , ] | 少妇, 姐姐, 美女, 少女, 嫂子 | 次子, 出生, 肾虚, 继任, 长兄 | |
| Primary set without 男孩/ 女孩 | 4 | [ , ] | 少妇, 女人, 女性, 美女, 少女 | 运气, 外围, 长兄, 离婚, 破解 | |
| Primary set without 男性/ 女性 | 4 | [ , ] | 少妇, 美女, 女人, 少女, 姐姐 | 运气, 长兄, 如何, 次子, 反而 | |
| Primary axis (5 pairs) | Extended axis (10 pairs) | |||||
|---|---|---|---|---|---|---|
| Specification | Estimate [95% CI] | Estimate [95% CI] | ||||
| No controls (Supplementary Fig. 4 ) | [ , ] | 2,924 | [ , ] | 2,924 | ||
| Excluding lexically gendered terms and seed words | [ , ] | 2,882 | [ , ] | 2,882 | ||
| Excluding the 5% most strongly gendered terms | [ , ] | 2,777 | [ , ] | 2,777 | ||
| Excluding the 10% most strongly gendered terms | [ , ] | 2,631 | [ , ] | 2,631 | ||
| , controlling for word frequency | [ , ] | 2,924 | [ , ] | 2,924 | ||
| Axis | Pairs | [95% CI] | Most feminine words | Most masculine words | |
| A. Axes built from several kinship/role pairs | |||||
| Primary set (5 pairs) | 5 | [ , ] | zan , seksi , pestān , bānu , set † | barādar , pedar , pesar , sardār , khān | |
| Extended set (10 pairs) | 10 | [ , ] | malake , zan , seksi , khānom , hāmele | mollā , Mohammad , hāji , āqā , khān | |
| Primary set without mard / zan | 4 | [ , ] | pestān , bānu , khānom , seksi , sadaf | barādar , pedar , shurid , farmāndehi , ben † | |
| Primary set without pesar / dokhtar | 4 | [ , ] | zan , seksi , ti † , gol , set † | barādar , mardi , sardār , pedar , jenāb | |
| Primary set without mardāne / zanāne | 4 | [ , ] | zan , mu , set † , ti † , seksi | barādar , pesar , pedar , shurid , mardi | |
| Language (gender system) | Norms | (coverage) | Held-out pairs | [95% CI], primary axis | [95% CI], extended axis | Leave-one-pair-out range | |
|---|---|---|---|---|---|---|---|
| Turkish (genderless) | Göz et al. 44 | 568 (100%) | 10/10 | [ , ] | [ , ] | to | |
| Italian (grammatical gender) | Montefinese et al. 79 | 1,114 (99%) | 10/10 | [ , ] | [ , ] | to | |
| French (grammatical gender) | Bonin et al. 15 | 1,633 (98%) | 10/10 | [ , ] | [ , ] | to | |
| Portuguese (grammatical gender) | Soares et al. 104 | 3,800 (100%) | 10/10 | [ , ] | [ , ] | to | |
| Indonesian (genderless) | Sianipar et al. 103 | 1,428 (96%) | 10/10 | [ , ] | [ , ] | to |
| Pooled (Fig. 1a) | Words only | Phrases only | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Model | 95% CI | 95% CI | 95% CI | ||||||
| FastText | [ , ] | 93,742 | [ , ] | 34,132 | [ , ] | 59,610 | |||
| word2vec | [ , ] | 85,229 | [ , ] | 33,963 | [ , ] | 51,266 | |||
| Wikipedia GloVe | [ , ] | 89,877 | [ , ] | 31,609 | [ , ] | 58,268 | |||
| ConceptNet | [ , ] | 97,080 | [ , ] | 38,341 | [ , ] | 58,739 | |||
| BERT | [ , ] | 102,842 | [ , ] | 39,953 | [ , ] | 62,889 | |||
| Words (Fig. 1c) | Phrases (averaged) | Words and phrases | |||||||
|---|---|---|---|---|---|---|---|---|---|
| Model | 95% CI | 95% CI | 95% CI | ||||||
| FastText | [ , ] | 33,938 | [ , ] | 58,821 | [ , ] | 92,759 | |||
| word2vec | [ , ] | 33,789 | [ , ] | 50,214 | [ , ] | 84,003 | |||
| Wikipedia GloVe | [ , ] | 31,446 | [ , ] | 57,487 | [ , ] | 88,933 | |||
| ConceptNet | [ , ] | 38,126 | [ , ] | 58,052 | [ , ] | 96,178 | |||
| BERT | [ , ] | 39,706 | [ , ] | 62,321 | † | [ , ] | 102,027 | ||
| Specification | Mean | 95% CI | Range | Negative | Sig. | per model |
| A. Lexical covariates | ||||||
| No controls (as in Fig. 1a; Supplementary Table 11 ) | [ , ] | to | 11/12 | 11 | 31,609–39,953 | |
| + word frequency | [ , ] | to | 10/12 | 9 | 31,609–39,953 | |
| + length (characters and syllables) | [ , ] | to | 10/12 | 10 | 31,609–39,953 | |
| + polysemy (number of WordNet senses) | [ , ] | to | 11/12 | 9 | 28,958–34,026 | |
| no controls, same words | [ , ] | to | 11/12 | 10 | 28,958–34,026 | |
| Specification | Mean | 95% CI | Range | Negative | Sig. | per model |
|---|---|---|---|---|---|---|
| No controls (as in Fig. 1a; Supplementary Table 11 ) | [ , ] | to | 12/12 | 12 | 51,266–62,889 | |
| + word frequency (mean of constituent words) | [ , ] | to | 12/12 | 12 | 50,839–61,308 | |
| + length (characters and number of words) | [ , ] | to | 12/12 | 12 | 51,266–62,889 | |
| + polysemy (WordNet senses of the phrase; indicator for a WordNet entry) | [ , ] | to | 12/12 | 12 | 51,266–62,889 | |
| + age of acquisition (mean of constituent words) 64 | [ , ] | to | 12/12 | 12 | 50,713–61,157 | |
| no controls, same phrases | [ , ] | to | 12/12 | 12 | 50,713–61,157 |
| Words | Phrases | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Model | None | + Freq. | + Length | + VAD | + All | None | + Freq. | + Length | + POS | + All |
| FastText | ||||||||||
| word2vec | ||||||||||
| Wikipedia GloVe | ||||||||||
| ConceptNet | † | |||||||||
| BERT | ||||||||||
| Words (12 models) | Phrases (12 models) | |||||
|---|---|---|---|---|---|---|
| Specification | Mean | 95% CI | Neg. (sig.) | Mean | 95% CI | Neg. (sig.) |
| All terms (as in Fig. 1a) | [ , ] | 11/12 (11) | [ , ] | 12/12 (12) | ||
| Excluding person nouns, lexically gendered terms, function words and names | [ , ] | 11/12 (10) | [ , ] | 12/12 (12) | ||
| As above, controlling for frequency, length and part of speech | [ , ] | 10/12 (8) | [ , ] | 10/12 (10) | ||
| Excluding the 5% most strongly gendered terms in each model | [ , ] | 10/12 (10) | [ , ] | 12/12 (12) | ||
| Excluding the 10% most strongly gendered terms in each model | [ , ] | 10/12 (10) | [ , ] | 12/12 (12) | ||
| Model | WordNet domain | words | [95% CI] | |
|---|---|---|---|---|
| Qwen3-Embedding-0.6B | verb.communication | 1,393 | [ , ] | |
| Qwen3-Embedding-0.6B | adj.all | 5,024 | [ , ] | |
| Qwen3-Embedding-0.6B | noun.artifact | 4,634 | [ , ] | |
| text-embedding-3-large | verb.communication | 1,393 | [ , ] | |
| text-embedding-3-large | adj.all | 5,024 | [ , ] | |
| text-embedding-3-large | noun.artifact | 4,634 | [ , ] |
| Assignment scheme | Words assigned | Domains | Abstract concrete (sign-test P ) | Median gap (s.d.) | Sig. | Sig. |
|---|---|---|---|---|---|---|
| Primary sense (main-text rule) | 22,831 | 37 | 32 ( ) | 0.10 | 17 | 0 |
| All senses (Supplementary Fig. 6 rule) | 39,430 | 40 | 35 ( ) | 0.12 | 20 | 0 |
| Domain-unambiguous words only | 13,481 | 22 | 19 ( ) | 0.12 | 8 | 0 |
| Random sense (200 replicates; min–max) | — | 36–38 | 30–37 ( ) | 0.10–0.15 | 13–23 | 0–1 |
| All terms (as in Fig. 3) a | Terms with 1 face | Terms with 1 face, adjusted for faces b | |
| Human judgements of concreteness (1–5; Fig. 3b) | |||
| Pearson’s | |||
| 95% CI | [ , ] | [ , ] | [ , ] |
| OLS slope (per scale point) | |||
| 95% CI (HC1) | [ , ] | [ , ] | [ , ] |
| P | |||
| Concreteness | Perceptual strength | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Specification | s.e. | / | P | s.e. | / | P | ||||
| (1) All terms; terms without faces set to zero (as in Fig. 3); no controls | 39,956 | 39,707 | ||||||||
| (2) Terms with 1 face; no controls | 35,913 | 35,677 | ||||||||
| (3) As (2), controlling for the number of faces (log) | 35,913 | 35,677 | ||||||||
| (4) As (3), plus word frequency (log) and classifier confidence | 35,913 | 35,677 | ||||||||
| (5) As (4), plus part-of-speech fixed effects | 35,913 | 35,677 | ||||||||
| Concreteness (human judgments) | Perceptual strength (Lancaster) | ||||||
| Specification | Estimate [95% CI] | terms | Estimate [95% CI] | terms | |||
| A. All search terms; terms without faces coded 0 (standardized ) | |||||||
| A1 | No controls (main-text figure) | [ , ] | 39,956 | [ , ] | 39,707 | ||
| A2 | + number of faces | [ , ] | 39,956 | [ , ] | 39,707 | ||
| A3 | + log(1 + faces) | [ , ] | 39,956 | [ , ] | 39,707 | ||
| A4 | + any-face indicator | [ , ] | 39,956 | [ , ] | 39,707 | ||
| Specification | 95% CI | slope | s.e. | P | ||||
| (a) Concreteness (1–5) | ||||||||
| (1) No controls | to | 36,124 | ||||||
| (2) Controlling for the number of images (log); as in Fig. 4 | to | 36,124 | ||||||
| (3) Controlling for the number of images (linear) | to | 36,124 | ||||||
| (4) Controlling for the number of images (linear and quadratic) | to | 36,124 | ||||||
| (5) Fixed effects for deciles of the number of images | to | 36,124 | ||||||
Gender Bias in Language Models
Gender bias in language models examines how systems encode and express gender stereotypes and how to evaluate and reduce their effects.
See all 75 papers