Inclusivity in Large Language Models: Personality Traits and Gender Bias in Scientific Abstracts

Applications, bias, and safety2024aircconline.comApproved editorial review

Authors: Naseela Pervez, Alexander J. Titus

Keywords: Computation and Language, Artificial Intelligence, Large Language Models, Gender Bias, Scientific Writing

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

2
Authors
20
Findings
32
Limitations
9
Evidence

Editorial summary

English

The study asks whether Claude 3 Opus, Gemini 1.5 Flash, and Mistral AI Large preserve LIWC patterns from human scientific abstracts when rewriting them, and whether differences between exclusively male- and female-authored texts reappear or grow. It starts from 3,390 randomly selected CORE abstracts: 418 are classified as female-only, 946 as male-only, and 2,026 as mixed-gender through `gender-extractor`, which infers gender from names. Each model receives the same zero-shot rewrite prompt with default parameters. LIWC-22 extracts 35 lexical, cognitive, affective, social, cultural, and motive indicators. For each indicator, Pearson correlation compares the original and rewritten values over 3,390 pairs; the gender analysis excludes the mixed majority and compares 946 against 418 through independent t-tests, separately for originals and each model. Published diagonal correlations are positive and range roughly from 0.22 to 0.92: word count correlates 0.35 for Claude, 0.80 for Gemini, and 0.86 for Mistral; discrepancy is 0.22 for all three, while anxiety, ethnicity, and technology often exceed 0.85. The paper reports p<0.05 for every non-NaN correlation and concludes that models preserve human textual traits. This only establishes covariation of LIWC counts in rewrites of the same content; it does not measure psychological personality, semantic equivalence, or lack of distortion. In the gender table, originals show, among others, t=-5.86 for word count, 3.93 for Tone, -6.74 for insight, -2.27 for cause, 3.82 for positive emotion, 5.70 for politeness, and 2.61 for risk. LLM outputs show new or clearer significant differences for achievement, conflict, and curiosity, while causation is no longer significant. Those shifts are descriptively interesting, but the study never directly tests amplification: it does not compare LLM-minus-original change between groups or estimate an interaction. The most serious error is interpreting t statistics as ratios: t=-5.86 becomes '5.86 times more words,' t=-6.74 becomes 'seven times more insight words,' and t=5.70 becomes 'five times more polite.' A t statistic is not a ratio of means; without means, dispersions, and effect sizes, no 'times more' statement follows. The analysis also lacks multiplicity correction, exact p values, intervals, and released data. The prompt contains neither gender nor author names: the model sees only the abstract, so an inherited difference may reflect preservation of topic or style rather than a model-introduced stereotype. The published version acknowledges class imbalance, disciplinary confounding, and absence of longitudinal analysis. Without the CORE subset, generations, code, exact model versions, and LIWC outputs, the work cannot be recomputed. The defensible contribution is an exploratory paired-LIWC comparison scheme; the evidence does not quantify bias magnitudes or establish that LLMs cause or amplify a gender gap.

Español

El estudio pregunta si Claude 3 Opus, Gemini 1.5 Flash y Mistral AI Large preservan en una reescritura los patrones LIWC de resúmenes científicos humanos y si las diferencias entre textos de autoría exclusivamente masculina o femenina reaparecen o se amplifican. Parte de 3.390 abstracts tomados aleatoriamente de CORE: 418 se clasifican como female-only, 946 como male-only y 2.026 como mixed-gender mediante la biblioteca `gender-extractor`, que infiere género a partir de nombres. Cada modelo recibe el mismo prompt zero-shot para reescribir el abstract con parámetros por defecto. LIWC-22 extrae 35 indicadores léxicos, cognitivos, afectivos, sociales, culturales y motivacionales. Para cada indicador se calcula la correlación de Pearson entre el valor original y el reescrito sobre los 3.390 pares; para el análisis de género se excluye la mayoría mixta y se comparan 946 frente a 418 mediante t-tests independientes, por separado en originales y en cada modelo. Las correlaciones diagonales publicadas son positivas y van aproximadamente de 0,22 a 0,92: word count correlaciona 0,35 con Claude, 0,80 con Gemini y 0,86 con Mistral; discrepancy queda en 0,22 en los tres, mientras anxiety, ethnicity y technology superan con frecuencia 0,85. El paper informa p<0,05 para todas las correlaciones no NaN y concluye que los modelos conservan rasgos del texto humano. Esa lectura debe limitarse a covariación de recuentos LIWC en reescrituras del mismo contenido: no mide personalidad psicológica, equivalencia semántica ni ausencia de distorsión. En la tabla de género, los originales presentan, entre otras, t=-5,86 en word count, 3,93 en Tone, -6,74 en insight, -2,27 en cause, 3,82 en positive emotion, 5,70 en politeness y 2,61 en risk. En salidas LLM aparecen diferencias significativas nuevas o más claras para achievement, conflict y curiosity, mientras la diferencia de causation ya no resulta significativa. Estos cambios son descriptivos y potencialmente relevantes, pero el estudio no prueba directamente una amplificación: no compara el cambio LLM-menos-original entre grupos ni estima una interacción. El error más grave es que el texto interpreta los estadísticos t como razones: t=-5,86 se convierte en “5,86 veces más palabras”, t=-6,74 en “siete veces más palabras de insight” y t=5,70 en “cinco veces más cortesía”. Un estadístico t no es un cociente de medias; sin medias, desviaciones y tamaños de efecto no puede sostenerse ninguna afirmación de “veces más”. Tampoco se corrigen las numerosas comparaciones ni se publican p exactos, intervalos o datos. El prompt no contiene género ni nombres: el modelo solo ve el abstract, de modo que una diferencia heredada puede ser mera conservación de tema o estilo, no un estereotipo introducido por el modelo. La versión publicada reconoce desequilibrio de clases, confusión por disciplina y falta de análisis temporal. Sin el subconjunto CORE, salidas, código, versiones exactas de los modelos y resultados LIWC, no es posible recalcular el trabajo. La contribución defendible es un esquema exploratorio para comparar perfiles LIWC emparejados; la evidencia no basta para cuantificar magnitudes de sesgo ni afirmar que los LLM causan o amplifican una brecha de género.

Research question

To what extent do three LLMs preserve the LIWC profile of scientific abstracts when rewriting them, and do they reproduce, reduce, or amplify the differences observed between texts classified as exclusively male or female authorship?

Method

3,390 abstracts from CORE are randomly sampled and the gender of their authors is inferred with `gender-extractor`. Claude 3 Opus, Gemini 1.5 Flash, and Mistral AI Large rewrite each text with a single zero-shot prompt and default parameters. LIWC-22 produces 35 indicators. For each indicator, the 3,390 original and rewritten values are correlated using Pearson. For gender, 2,026 mixed-authorship works are excluded and two-sample t-tests are run between 946 male-only and 418 female-only on originals and on each LLM set. The editorial audit read and rendered the 16-page publication, compared it with arXiv v1, verified tables, DOI, and venue, reviewed the statistical interpretation, and searched without finding associated code or data.

Sample: 3,390 abstracts: 418 classified as written only by women, 946 only by men, and 2,026 by mixed teams. The correlation uses the 3,390 pairs per model; the t-tests use 1,364 originals and their rewrites, excluding the 59.8% mixed. If all abstracts were processed as described, the design comprises 3,390 originals and 10,170 rewrites, but the texts, identifiers, and matrices are not published.

Findings

  • The published version appears in NLAI 2024 / CS & IT 14(22) with DOI 10.5121/csit.2024.142201 and adds limitations absent from arXiv v1.
  • The sample is distributed across 418 female-only, 946 male-only, and 2,026 mixed-gender.
  • The diagonal LIWC original-rewrite correlations are positive between 0.22 and 0.92, except Segment NaN.
  • Word count correlates 0.35 with Claude, 0.80 with Gemini, and 0.86 with Mistral.
  • Discrepancy is the least aligned indicator, r=0.22 across all three models.
  • Ethnicity reaches 0.86/0.92/0.91 and technology 0.86/0.91/0.89 for Claude/Gemini/Mistral.
  • The paper declares p<0.05 for all non-NaN correlations, without correction for multiplicity.
  • On originals, t=-5.86 for word count indicates a standardized difference, not a 5.86x ratio.
  • On originals, insight t=-6.74, cause t=-2.27, and differ t=-3.20 show differences with a negative sign depending on the group order used.
  • Tone t=3.93, positive tone t=3.88, and positive emotion t=3.82 have a positive sign on originals.
  • Politeness t=5.70 and risk t=2.61 are significant on originals.
  • Achievement is not significant on originals (t=-0.78) but is significant in Claude, Gemini, and Mistral (-3.22/-2.15/-2.67).
  • Curiosity is not significant on originals (1.06) but is significant across all three LLMs (3.75/2.60/2.66).
  • Conflict stands at -1.95 on originals and shifts to -2.53/-2.58/-2.00 across the three LLMs.
  • Cause is significant on originals (-2.27) but not in any rewrite (-0.51/-1.56/-1.38).
  • The Tone difference is reduced in magnitude in Claude, Gemini, and Mistral (1.77/2.37/2.43) compared with the human 3.93.
  • The paper repeatedly interprets |t| as 'times more', a mathematically invalid inference.
  • No means, standard deviations, degrees of freedom, exact p-values, Cohen's d, or confidence intervals are published to allow magnitude estimation.
  • There is no interaction test or comparison of the rewritten-minus-original change between groups.
  • No repository, dataset, outputs, API code, or public LIWC matrices associated with the study were found.

Limitations

  • gender-extractor infers gender from names and does not measure self-identified gender identity.
  • No version, cultural coverage, threshold, unknown rates, or validation of name inference are reported.
  • The scheme reduces gender to male/female and excludes non-binary identities.
  • An abstract with multiple authors is attributed to a single collective category without modeling author number, order, or contribution.
  • 2,026 mixed-authorship works, 59.8% of the sample, are excluded from the gender analysis.
  • The groups are imbalanced, 946 versus 418; the paper acknowledges this but does not adjust.
  • Discipline, topic, and jargon may simultaneously explain inferred gender and LIWC; there are no field controls.
  • Publication year is not controlled, another limitation acknowledged in the published version.
  • Sampling is described as random without a sampling frame, seed, date, language, or inclusion criteria.
  • CORE identifiers are not published, so duplicates, fields, or attributions cannot be audited.
  • The models lack exact IDs, query dates, provider/deployment, and API versions.
  • Using default parameters prevents reproducing temperature, top-p, length, and the service's changing behavior.
  • Only one prompt is used and sensitivity to alternative instructions is not evaluated.
  • LIWC-22 is proprietary and the exact execution configuration or dictionary is not released.
  • LIWC counts describe word patterns; calling them author personality exceeds what is measured.
  • A positive correlation does not imply equality of level, low distortion, or semantic fidelity.
  • Correlating rewrites of the same content favors alignment through shared topic and vocabulary.
  • With n=3,390, p<0.05 significance can coexist with weak associations such as r=0.22.
  • Dozens of correlations and at least 35 t-tests per corpus are tested without correction for multiple comparisons.
  • The t-test version, equal-variance assumption, distribution diagnostics, or robustness are not specified.
  • The t-statistics are misinterpreted as ratios between means.
  • There are no group means or dispersions to verify direction and magnitude.
  • Comparing which tests are significant does not equate to testing that two effects differ.
  • To demonstrate amplification, an interaction or a comparison of gender-paired changes would be required.
  • The prompt does not include gender or names, so the model does not directly receive the audited category.
  • Differences in outputs may be inherited from the original content rather than introduced by the model.
  • Factual accuracy, meaning preservation, quality, plagiarism, or changes in conclusions are not evaluated.
  • There is no human evaluation of style, inclusivity, or perceived harm.
  • Only abstracts are studied; findings do not generalize to full manuscripts or real review processes.
  • The Mistral AI Large reference points to the Mistral 7B work and does not document the commercial model used.
  • The absence of data, code, and outputs prevents reproducing any coefficient or t-statistic.
  • The study does not measure citations, acceptance, visibility, or consequences for real authors.

What the study does not establish

  • It does not demonstrate the psychological personality of authors or models from LIWC.
  • It does not establish that a t-value of 5.86 means 5.86 times more of a characteristic.
  • It does not prove that LLMs know or correctly infer the gender of the authors.
  • It does not isolate gender from discipline, topic, period, language, or team composition.
  • It does not demonstrate amplification because it does not directly test changes in effect between original and rewrite.
  • It does not establish LLM causality when differences may come from the input text.
  • It does not quantify effect sizes or the practical relevance of the differences.
  • It does not prove that a high LIWC correlation preserves meaning or scientific quality.
  • It does not generalize to non-binary identities, mixed teams, other tasks, or current models.
  • It does not demonstrate effects on publication, citations, career, inclusion, or real discrimination.

Traceability

Scope: Full text

Version: NLAI 2024 / Computer Science & Information Technology vol. 14 no. 22, pages 1–16; DOI 10.5121/csit.2024.142201

Consulted source: https://aircconline.com/csit/papers/vol14/csit142201.pdf

Review: Codex published/full-text, visual and statistical-validity audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Claude 3 Opus, exact snapshot and API date not reported
  • Gemini 1.5 Flash, exact snapshot and API date not reported
  • Mistral AI Large, exact snapshot and API date not reported

Instruments and metrics

  • LIWC-22 proprietary dictionaries
  • 35 selected LIWC lexical, cognitive, affective, social, cultural and motive indicators
  • Pearson correlation on paired original/rewrite feature values
  • Independent two-sample t-tests for male-only versus female-only groups
  • gender-extractor name-based gender inference
  • Single zero-shot abstract-rewriting prompt

Data used

  • Unreleased random subset of 3,390 scientific abstracts from CORE
  • Unreleased 3,390 rewrites from each of Claude, Gemini and Mistral
  • Unreleased LIWC-22 feature matrices for original and rewritten texts
  • Published Table 2 correlation coefficients and Table 3 t statistics
  • arXiv:2406.19497v1 preserved as the earlier manuscript version

Evidence and location

  • Questions, motivation, and LIWC framework: Published paper pp. 1–5, Introduction, Related Work and Sections 3.1–3.3
  • CORE sample and gender assignment: Published paper p. 4, Section 3.1 and Table 1
  • Models, prompt, and default parameters: Published paper p. 5, Section 3.2
  • Correlations by indicator: Published paper pp. 6–9, Section 4.1, Table 2 and Figures 2–3
  • t-tests and 'times more' claims: Published paper pp. 9–12, Section 4.2, Table 3 and Figure 4
  • Imbalance, discipline, time, and intersectionality: Published paper pp. 12–13, Limitations and Future Works
  • Venue, DOI, date, and published version: AIRCC CS & IT volume 14 number 22 record and published PDF, checked 15 Jul 2026
  • Previous version and editorial differences: arXiv:2406.19497v1 full text compared with November 2024 publication
  • Absence of reproducible artifact: Publisher/arXiv associated-material records and GitHub title/author search, checked 15 Jul 2026