Theory-Grounded Measurement of U.S. Social Stereotypes in English Language Models

Applications, bias, and safety2022ACL AnthologyApproved editorial review

Authors: Yang Trista Cao, Anna Sotnikova, Hal Daumé III, Rachel Rudinger, Linda Zou

Keywords: Natural Language Processing, Social Stereotypes, Language Models, Intersectional Identities, ABC Stereotype Model

Source: Open primary source (opens in a new tab)

5
Authors
20
Findings
25
Limitations
10
Evidence

Editorial summary

English

Cao and coauthors try to move language-model stereotype measurement away from opportunistic lists of occupations or adjectives. They adopt the Agency-Beliefs-Communion (ABC) model from social psychology, organizing group perception into 16 bipolar trait pairs and three dimensions: agency or socioeconomic success, conservative-progressive beliefs, and communion or warmth. Pairs include powerless–powerful, poor–wealthy, traditional–modern, religious–science-oriented, cold–warm, and threatening–benevolent. The framework gains coverage and comparability across groups at the cost of specificity: it can identify an abstract association with threatening or dominant, not a concrete trope such as the angry Black woman stereotype.

The paper compares three measurement families on BERT-large-cased and RoBERTa-large. ILPS asks how much a group raises the probability of an adjective in a masked template and normalizes against a masked group; ILPS* extends the calculation to words split into multiple subwords. CEAT measures contextualized embedding distances with up to 1,000 Reddit sentences per word. The new Sensitivity Test (SeT) asks how much the model's final linear matrix must be perturbed to make a trait the highest-logit prediction by a margin of one, normalized against a group-free template. SeT is neither an observed probability nor an estimate of training frequency. It is a local output-layer robustness measure under a specific mathematical intervention.

The human reference comes from an IRB-approved survey. All 247 Prolific participants were paid $2 for roughly ten minutes, $12 per hour, slightly below the cited Maryland minimum wage of $12.20, and 133 passed three checks: recalling the group, recalling a high- and low-rated trait, and achieving at least 80% consistency on a repeated fifth group. Each participant rated four unique groups and one repeat; the analysis targets 20 valid annotations for each of 25 groups and 16 scales. The question asks how the participant thinks US society views a group, not for their personal opinion, using a 0–100 slider. Participants cover 26 states, but 63.3% live in California, New York, Texas, or Florida, and more than 96% are at most 40 years old.

Templates are not fixed before evaluation. For each model-measure combination, the authors test variants such as That [group] is [trait], All [group] are [trait], and [group] should be [trait], then select up to two that maximize Kendall's τ in a pilot using four groups, Asian, Black, Hispanic, and immigrant, with five annotations per group. Across all 25 groups, RoBERTa generally outperforms BERT. RoBERTa-SeT has the largest overall τ, 0.199, and P@3 of 0.653; RoBERTa-ILPS* reaches 0.175 and the same 0.653, while RoBERTa-ILPS reaches 0.169 and 0.620. These are moderate associations, not equivalence between model and population. The overall result also reuses the pilot groups. In the appendix check that excludes them, SeT obtains τ=0.174 and ILPS* τ=0.173, reducing the new measure's advantage to 0.001. P@3 dichotomizes each human mean at 50 and counts overlap among the three highest and lowest traits; agreement on roughly two of three does not itself express intensity, uncertainty, or interpersonal agreement.

For compound identities, the study applies RoBERTa-SeT to logically admissible pairs of individual groups and retains both orders where grammar permits. Mean correlation between a pair and its most similar component is 0.56; correlations with the first and second components are 0.43 and 0.46, and between reversed orders 0.69. Order appears less consequential than the presence of a dominant component, but no equivalence test or intervals are reported. A rule requiring at least a 0.1 correlation difference labels age and politics as dominant domains and race and nationality as dominated. To identify emergent traits, the authors subtract the maximum component score from the pair score. Examples include Hispanic unemployed–egotistic, Democrat teenager–altruistic, and male doctor–benevolent, but the paper acknowledges that many cases merely lift a component with a very low score toward the average. Against Ghavami and Peplau's race-gender intersection traits, detection reaches precision 0.83 and recall 0.65; a random baseline already reaches 0.72 and 0.50, and no confidence interval or difference test is supplied.

The public MIT-licensed artifact preserves code, a datasheet, and three JSON files, but it does not reproduce the paper end to end. It includes 26 aggregate groups, the average people control explains the difference from the 25 analyzed groups, 133 response records, and 134 demographic records whose identifiers were deliberately delinked. Of the 133 response records, 123 contain four groups and ten contain only three; one Jewish people annotation contains one trait rather than 16. Aggregate means match released microdata apart from rounding and two women cells: poor–wealthy is 57.1 in the released responses versus 56.4 in the aggregate, while untrustworthy–trustworthy is 57.1 versus 57.95. The repeated raw ratings and preprocessing code are absent, so those differences cannot be reconciled.

Code-paper divergences are more consequential. The PDF equation defines SeT through the logarithm of a distance ratio, but quick_set.py computes log(1-‖W-W'‖) for group and prior and subtracts the two; for multi-subword traits it retains the minimum, while the text says it takes the maximum. The paper says that after a synonym experiment it reverted to the 32 ABC adjectives; test_templates.py uses two to four synonyms per pole and takes their median, including misspellings such as athiestic and avante-garde. The script is fixed to RoBERTa and calls only SeT, does not save the tables promised by the README, and assumes a missing results directory. There is no pinned environment, released output, template-selection analysis, correlation or P@3 code, or intersectional code. The Python syntax compiles, but the reported numbers cannot be regenerated from the twelve deposited files.

The defensible contribution is conceptual: a shared psychological framework enables broad group comparisons and shows that BERT/RoBERTa associations align only moderately, and very template-dependently, with an aggregate US human reference. SeT is an interesting local-robustness hypothesis, not an already validated superior measure: on groups withheld from template selection it is effectively tied with ILPS*. The study also does not show that models understand identities or that their scores cause discrimination in deployment. A reference-quality benchmark would need a fixed implementation, published outputs and analysis, clean selection/test separation, uncertainty estimates, and renewed models and cultural samples.

Español

Cao y coautores intentan que la medición de estereotipos en modelos de lenguaje deje de depender de listas oportunistas de profesiones o adjetivos. Adoptan de la psicología social el modelo Agency-Beliefs-Communion (ABC), que organiza la percepción de grupos en 16 pares bipolares y tres dimensiones: agencia o éxito socioeconómico, creencias conservadoras-progresistas y comunión o calidez. Los pares incluyen powerless–powerful, poor–wealthy, traditional–modern, religious–science-oriented, cold–warm y threatening–benevolent. El marco gana cobertura y comparabilidad entre grupos, pero pierde especificidad: detecta una asociación abstracta con threatening o dominant, no tropos concretos como angry Black woman.

El artículo compara tres familias de medida sobre BERT-large-cased y RoBERTa-large. ILPS observa cuánto eleva un grupo la probabilidad de un adjetivo en una plantilla enmascarada y la normaliza con un grupo enmascarado; ILPS* extiende el cálculo a palabras divididas en varios subwords. CEAT mide distancias de embeddings contextualizados usando hasta 1.000 frases de Reddit por palabra. La contribución nueva, Sensitivity Test (SeT), pregunta cuánto habría que perturbar la última matriz lineal del modelo para que un rasgo pase a ser la predicción con mayor logit por un margen de uno, y normaliza esa distancia respecto de una plantilla sin grupo. SeT no mide probabilidad observada ni frecuencia de entrenamiento; mide una robustez local de la capa de salida bajo una intervención matemática específica.

La referencia humana procede de una encuesta aprobada por IRB. Se pagaron $2 a 247 participantes de Prolific por unos diez minutos, $12/h, ligeramente por debajo del salario mínimo de Maryland citado, $12,20, y 133 superaron tres controles: recordar el grupo, recordar rasgos marcados alto y bajo, y alcanzar al menos 80% de consistencia al repetir el quinto grupo. Cada participante calificó cuatro grupos únicos y una repetición; el análisis busca 20 evaluaciones válidas para cada uno de 25 grupos y 16 escalas. La pregunta pide cómo cree la persona que la sociedad estadounidense ve al grupo, no su opinión personal, en un slider 0–100. La muestra cubre 26 estados, pero 63,3% reside en California, Nueva York, Texas o Florida y más de 96% tiene 40 años o menos.

Las plantillas no se fijan antes de evaluar. Para cada combinación de modelo y medida, los autores prueban variaciones como That [group] is [trait], All [group] are [trait] o [group] should be [trait], y eligen hasta dos que maximizan Kendall τ en un piloto de cuatro grupos, Asian, Black, Hispanic e immigrant, con cinco anotaciones por grupo. Sobre los 25 grupos, RoBERTa supera en general a BERT. RoBERTa-SeT obtiene el mayor τ global, 0,199, y P@3 de 0,653; RoBERTa-ILPS* queda en 0,175 y también 0,653, mientras RoBERTa-ILPS alcanza 0,169 y 0,620. Son asociaciones moderadas, no una equivalencia entre modelo y población. Además, el resultado global reutiliza los grupos del piloto. En el control del apéndice que excluye esos cuatro grupos, SeT da τ=0,174 e ILPS* τ=0,173: la ventaja de la nueva medida se reduce a 0,001. P@3 binariza cada promedio humano en 50 y cuenta coincidencias entre los tres rasgos superiores e inferiores; que coincidan aproximadamente dos de tres no informa por sí solo de intensidad, incertidumbre o acuerdo entre personas.

Para identidades compuestas, el trabajo aplica RoBERTa-SeT a pares lógicamente admisibles de los grupos individuales y conserva ambos órdenes cuando la gramática lo permite. La correlación media del par con su componente más parecido es 0,56; con el primer y segundo componente es 0,43 y 0,46, y entre órdenes invertidos 0,69. El orden parece importar menos que la presencia de un componente dominante, aunque no se presenta un test de equivalencia ni intervalos. Una regla con diferencia de correlación de al menos 0,1 clasifica edad y política como dominios dominantes, y raza y nacionalidad como dominados. Para buscar rasgos emergentes, los autores restan al score del par el máximo de sus dos componentes. Encuentran asociaciones como Hispanic unemployed–egotistic, Democrat teenager–altruistic y male doctor–benevolent, pero admiten que muchos casos solo elevan a la media un componente con score muy bajo. Frente a rasgos de intersecciones raza-género de Ghavami y Peplau, el detector alcanza precisión 0,83 y recall 0,65; la referencia aleatoria ya obtiene 0,72 y 0,50, y no se ofrecen intervalos ni prueba de diferencia.

El artefacto público, bajo licencia MIT, conserva código, datasheet y tres JSON, pero no permite reproducir el artículo de extremo a extremo. Incluye 26 grupos agregados, el control average people explica la diferencia con los 25 analizados, 133 registros de respuestas y 134 registros demográficos cuyos identificadores se desligaron deliberadamente. De los 133 registros de respuesta, 123 contienen cuatro grupos y diez solo tres; una anotación de Jewish people tiene un único rasgo en vez de 16. Los promedios agregados coinciden con los microdatos salvo redondeo y dos celdas de women: poor–wealthy es 57,1 en los datos liberados frente a 56,4 en el agregado, y untrustworthy–trustworthy 57,1 frente a 57,95. No hay datos crudos de la repetición, código de limpieza ni forma de reconciliar esas diferencias.

Las divergencias del código son más importantes. La ecuación del PDF define SeT mediante el logaritmo de una razón de distancias, pero quick_set.py calcula log(1-‖W-W'‖) para grupo y prior y luego los resta; para rasgos con varios subwords conserva el mínimo, mientras el texto dice tomar el máximo. El artículo afirma que, tras un preliminar con sinónimos, volvió a los 32 adjetivos ABC; test_templates.py usa de dos a cuatro sinónimos por polo y toma la mediana, con errores como athiestic y avante-garde. El script está fijado a RoBERTa y solo llama a SeT, no guarda las tablas prometidas por el README y presupone un directorio results ausente. Tampoco se publican dependencias fijadas, outputs, selección de plantillas, correlaciones, P@3 ni código interseccional. La sintaxis compila, pero los números del trabajo no pueden regenerarse con los doce archivos depositados.

La contribución defendible es conceptual: un marco psicológico común permite comparar muchos grupos y muestra que las asociaciones de BERT/RoBERTa se alinean solo de forma moderada y muy dependiente de plantilla con un promedio humano estadounidense. SeT es una hipótesis interesante de robustez local, no una medida ya validada como superior: en grupos no usados para seleccionar plantillas empata prácticamente con ILPS*. El estudio tampoco demuestra que el modelo comprenda identidades ni que sus scores produzcan discriminación en un sistema desplegado. Para usarlo como benchmark de referencia habría que fijar la implementación exacta, publicar outputs y análisis, separar selección y test, informar incertidumbre y renovar tanto modelos como muestra cultural.

Research question

To what extent do the group-trait associations of BERT and RoBERTa, measured with CEAT, ILPS, ILPS* or the new SeT and organized through the ABC model, coincide with aggregated judgments from U.S. participants; and what patterns appear when combining identities?

Method

Comparison of intrinsic measures and human survey. 32 poles of 16 ABC pairs are scored for more than 50 individual groups using BERT-large-cased and RoBERTa-large. CEAT uses contextual embeddings; ILPS masked probabilities; ILPS* corrects subwords; SeT computes a minimal perturbation of the output layer. Templates are chosen on four pilot groups and compared with human averages of 25 groups using Kendall τ and P@3. RoBERTa-SeT is then applied to pairs of identities to study order, domain dominance and emergent traits. The editorial audit read and rendered the 20 pages, verified ACL, froze the public commit, inspected code and datasheet, compiled the scripts and reconciled the aggregated JSON with the released microdata.

Sample: 247 U.S. participants from Prolific completed the survey and were paid; 133 passed the controls and feed the analysis. Each scored four unique groups and repeated a fifth, 16 pairs per group; 20 valid annotations were sought for each of 25 groups across five domains. The sample comes from 26 states, 63.3% from four states, more than 96% are 40 years old or younger, and the article reports 48.2% men, 49.6% women and 2.2% genderqueer, agender or questioning. The released demographic JSON contains 134, not 133, records.

Findings

  • RoBERTa generally shows greater alignment with human averages than BERT, but the result depends on measure and template.
  • Across the 25 groups, RoBERTa-SeT obtains the highest Kendall τ, 0.199, versus 0.175 for ILPS*, 0.169 for ILPS and 0.019 for CEAT.
  • RoBERTa-SeT and ILPS* tie at P@3=0.653; ILPS reaches 0.620 and CEAT 0.500.
  • BERT-SeT obtains τ=0.116 and P@3=0.613; the architecture change alters both alignment and preferred templates.
  • Overall significant correlations remain moderate and vary widely by group; some are negative.
  • Excluding Asian, Black, Hispanic and immigrant, used to select templates, SeT and ILPS* on RoBERTa remain nearly equal: τ=0.174 and 0.173.
  • RoBERTa tends to prefer templates such as That [group] is [trait], while BERT obtains better results with All [group] are [trait] or [group] should be [trait].
  • CEAT does not produce scores for some groups absent from the Reddit 2014 corpus, including Hispanic, non-binary and autistic in the per-group tables.
  • The mean correlation of a compound identity with the component it most resembles is 0.56.
  • Mean correlations with the first and second component are 0.43 and 0.46; across inverted orders, 0.69.
  • With a descriptive threshold of difference 0.1, age and politics dominate other domains, while race and nationality are usually dominated.
  • The race-gender emergent trait detector obtains precision 0.83 and recall 0.65 versus 0.72 and 0.50 for a random baseline.
  • Inspection of emergent examples themselves shows mixed face validity and effects of elevating components with initially very low scores.
  • Only 133 of 247 participants pass the three quality controls; 114 are excluded although they are paid.
  • Human averages conceal disagreement, intensity and different prototypes: 50 may mean neutral consensus or a polarized mix of 0 and 100.
  • The artifact includes data and an MIT license, but no outputs or scripts for the correlations, P@3, template selection or intersectional analysis.
  • Two aggregated cells for women are not the mean of the published microdata: poor-wealthy 56.4 versus 57.1 and untrustworthy-trustworthy 57.95 versus 57.1.
  • Ten of 133 complete records contain only three groups and one Jewish people annotation contains one trait; the processing that produced the aggregate is not published.
  • The published SeT code uses log(1-distance) and minimum across subwords, instead of literally implementing the distance ratio and maximum described in the PDF.
  • The main script uses synonyms and median despite the final method declaring a return to the 32 canonical adjectives.

Limitations

  • Only two masked language models prior to 2020 are evaluated; no autoregressive generators, instructed models or current systems.
  • The survey and interpretation are restricted to English and U.S. stereotypes.
  • The sample is young and geographically concentrated, and does not probabilistically represent the U.S. population.
  • The exclusion rate is 114 of 247, so those who pass memory and consistency tests may be a selective subset.
  • Asking what society thinks reduces direct opinion but introduces attribution of consensus, feeling unqualified and possible reinforcement of stereotypes.
  • Averaging judgments confuses consensual neutrality with polarization and erases the distribution across participants.
  • The defaulting effect may make man implicitly mean cis, heterosexual and White for both humans and corpora.
  • ABC traits offer abstract coverage, but do not capture concrete stereotypes, context, pragmatics or natural deployment language.
  • Templates are artificial and some, such as All [group] are or should be, express stronger generalizations than ordinary queries.
  • Template selection uses human responses and four groups that reappear in the global evaluation, generating optimism from fit to the pilot.
  • The test-only advantage of SeT over ILPS* is 0.001 and is not statistically tested as a difference between dependent correlations.
  • P@3 binarizes at 50 and provides no intervals, significance, calibration or prevalence-adjusted baseline for poles.
  • CEAT depends on the group name appearing in Reddit 2014 and is missing for several groups, so comparisons do not have identical coverage.
  • For subwords, SeT only computes a bound based on individual pieces; the semantic composition of the full term is not resolved.
  • SeT intervenes only on the last linear layer and does not characterize the robustness of internal representations or full generative behavior.
  • The dominance rule uses a 0.1 threshold without inferential justification and summarizes pair distributions in a single mean.
  • The conclusion that order does not matter lacks an equivalence test, intervals and per-pair distribution.
  • Emergence detection by difference from the maximum may confuse genuine interaction with regression to the mean or unstable scales.
  • The random baseline precision of 0.72 is already high; the improvement to 0.83 is not accompanied by uncertainty or a significance test.
  • The manual correspondence between external traits and ABC dimensions is not published as a reproducible table nor is its sensitivity analyzed.
  • The demographic JSON has 134 records and the response one 133; their identifiers are deliberately unlinked, preventing reproduction of tables by subgroup.
  • The microdata contain incomplete records and two non-trivial discrepancies with the aggregate; there is no raw data or preprocessing code.
  • There are no requirements, lockfile, exact versions of Transformers/PyTorch, tests, CI, outputs or reproducible command.
  • The public code and the PDF diverge on the SeT distance transformation, subword aggregation and use of synonyms.
  • The study measures intrinsic associations; it does not verify their transfer to classification, completion, recommendation or decisions about people.

What the study does not establish

  • It does not demonstrate that BERT or RoBERTa have beliefs, intentions or human psychology.
  • It does not validate SeT as an unequivocally superior measure to ILPS*; outside the pilot their correlations are practically identical.
  • It does not prove that intrinsic associations cause discrimination or harm in a deployed system.
  • It does not establish that the average of 133 participants is the truth about U.S. stereotypes.
  • It does not demonstrate that each emergent association is a socially documented intersectional stereotype.
  • It does not show that age or politics dominate identities in people; it only describes RoBERTa-SeT scores under its rule.
  • It does not generalize to other languages, countries, periods, models or natural formulations.
  • It does not allow regenerating the tables and conclusions with the public repository without reconstructing code, versions, outputs and absent decisions.
  • It does not authorize using the data for profiling, classifiers or real decisions about members of the studied groups.

Traceability

Scope: Full text

Version: NAACL 2022 main paper 92, pages 1276–1295; DOI 10.18653/v1/2022.naacl-main.92

Consulted source: https://aclanthology.org/2022.naacl-main.92.pdf

Review: Codex full-text, visual, code and released-data consistency audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • bert-large-cased masked language model
  • roberta-large masked language model

Instruments and metrics

  • Agency-Beliefs-Communion model with 16 bipolar trait pairs and 32 canonical adjectives
  • Sensitivity Test (SeT) with margin gamma=1 and column-squishing update
  • Increased Log Probability Score (ILPS)
  • Multi-subword modified ILPS (ILPS*)
  • Contextualized Embedding Association Test (CEAT)
  • Template bank spanning number, interrogative, adverbial, belief, expectation, word order and comparison variants
  • Kendall rank correlation
  • Precision at 3 for positive and negative trait poles
  • Prolific/Qualtrics 0–100 group-trait survey with three quality checks

Data used

  • Crowdsourced U.S. Stereotypes Measured across Agency, Beliefs, and Communion
  • Aggregated_scores.json with 26 groups by 16 trait pairs
  • data_full.json with 133 retained response records
  • Annotators_demographics.json with 134 delinked demographic records
  • Reddit 2014 contextual sentences used by the adopted CEAT method
  • Ghavami and Peplau 2013 intersectional stereotype traits for external comparison
  • GitHub TristaCao/U.S_Stereotypes at commit 5b1f382d838c6aff4c392dc60bd23c1a67e54949

Evidence and location

  • ABC framework, coverage, advantages and trade-off with specificity: NAACL paper pp. 1-3, Introduction, Background and Tables 1-3
  • Definition of ILPS, CEAT, SeT, subwords and templates: NAACL paper pp. 3-5, Section 3, equations and Table 4
  • Survey, payment, controls, participants and demographics: NAACL paper pp. 5-6, Section 4; Appendix D-E and Tables A14-A18
  • Template selection and global/test-only alignment: NAACL paper pp. 6-7, Section 5.1, Tables 5-6; Appendix Tables A11-A13 and A19
  • Order, dominance and emergent traits in pairs: NAACL paper pp. 7-9, Section 5.2; Appendix Tables A8-A10
  • Limitations, risk of reinforcement and non-deployed scope: NAACL paper pp. 9-10, Sections 6-7
  • Metadata, authorship, DOI and pagination: ACL Anthology record 2022.naacl-main.92, checked 15 Jul 2026
  • Composition, license and dataset restrictions: GitHub TristaCao/U.S_Stereotypes commit 5b1f382, LICENSE and dataset/Datasheet.md
  • Incomplete records and aggregate-microdata discrepancies: Editorial reconciliation of Aggregated_scores.json, data_full.json and Annotators_demographics.json, 15 Jul 2026
  • Divergences between equation and code, synonyms and lack of reproducible analysis: Repository quick_set.py, test_templates.py, code/readme.md and complete file inventory, audited 15 Jul 2026