Your browser does not fully support modern features. Please upgrade for a smoother experience.
Please note this is a comparison between Version 1 by Clotilde Cicatiello and Version 2 by Perry Fu.
Gender bias in generative artificial intelligence (GenAI) is both a technical and a social phenomenon: it emerges from historically patterned data, model design, and interactions in institutional use, and it cannot be understood by engineering or by social critique alone. This critical integrative review develops a more differentiated account. It connects feminist epistemology, Science and Technology Studies, critical AI scholarship, natural language processing, and governance research to examine five levels: historical knowledge production, technical representation and generation, benchmark evaluation, institutional deployment, and accountability. The review explains tokenization, next-token prediction, transformers, and the transition from static embeddings to contemporary language models before assessing evidence from standard fairness tests—coreference tests (WinoBias), sentence-pair tests (CrowS-Pairs), and stereotype tests (StereoSet)—as well as open-ended generation, multilingual testing, and text-to-image systems. It shows that measured bias varies with task, prompt, language, model version, and metric. What a test records and what that record means are therefore distinct questions: measurements are situated and depend on the instrument, and their interpretation draws on theory rather than following from the numbers alone. Evidence from employment, education, healthcare, and translation further indicates that the relevant unit of analysis is the model-in-context—the model together with the institution and workflow in which its outputs are used. Technical mitigation can reduce specific harms but does not repair unequal criteria, incomplete evidence bases, or weak institutional accountability. The review proposes a multilevel governance approach combining technical evaluation, documentation, professional and community oversight, appeals, remedies, and public-interest knowledge infrastructure. Its distinctive contribution is to connect three observations usually kept apart—how bias is measured, how generative systems concentrate epistemic authority, and how statistical learning is oriented toward past data—and to show why democratic and feminist governance can keep alternative technological futures open.
- generative artificial intelligence
- gender bias
- large language models
- feminist epistemology
- algorithmic fairness
- benchmarking
- governance
- epistemic justice
Generative artificial intelligence (GenAI) has moved rapidly from experimental research into writing assistance, image production, education, clinical documentation, recruitment, and public administration. This diffusion has sharpened a longstanding question in the study of technology—one better posed as both/and than as either/or: how computational systems at once reflect existing social inequalities and reorganise and extend them. Gender is a particularly revealing case because it operates simultaneously as a category of representation, an organizing principle of institutions, a source of historical data imbalance, and a dimension of unequal exposure to technological benefits and harms [1][2][3][4][5][6][7][8][9][10][1,2,3,4,5,6,7,8,9,10].
Research on bias in natural language processing (NLP) has produced substantial evidence that statistical representations can encode stereotypical associations, that generated text can reproduce occupational and interpersonal expectations, and that model behaviour varies across prompts, languages, tasks, and alignment procedures [11][12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27][28][29][30][31][32][33][34][35][36][37][38][39][11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39]. However, the literature also warns against treating every difference in an output as proof of discrimination or assuming that a low score on a benchmark guarantees equitable performance in practice. Bias is not a single measurable property. It may concern representational harm, unequal allocation, reduced service quality, epistemic exclusion, or the transfer of authority from accountable professionals to opaque systems [23][30][31][32][33][23,30,31,32,33].
Existing surveys provide increasingly sophisticated taxonomies of metrics, datasets, and technical mitigation strategies [30]. The distinctive contribution of this review is different. It develops a critical integrative review and conceptual analysis that connects five levels often studied separately: the historical production of gendered knowledge; the technical mechanisms of contemporary generative models; the limits of benchmark-based evaluation; the institutional contexts in which models are deployed; and the governance choices that determine who can contest, correct, or benefit from these systems. The review therefore does not present new experimental findings and does not claim that all models or deployments produce identical effects.
The central argument is that gender bias in GenAI should be understood as a sociotechnical relation rather than as a defect located only in data or model weights. Technical interventions can reduce specific measured associations and are therefore valuable, but their effects depend on the metric, the task, the model, and the institution in which the model is used [19][20][31][32][37][19,20,31,32,37]. Conversely, structural critique is incomplete when it treats mitigation as irrelevant or technologically impossible. A balanced account must distinguish what empirical studies demonstrate from the interpretations offered by feminist epistemology, critical AI studies, digital sociology, and decolonial theory.
In this review, gender bias does not name a single quantity but a family of patterned asymmetries in how generative systems represent people, treat them, and distribute attention and resources across gendered positions—asymmetries produced socially before they are learned statistically. Following feminist epistemology, gender is treated here as a relation of power and knowledge rather than a fixed attribute of individuals, and three registers in which bias becomes legible are distinguished: representational (who is depicted, how, and as what), allocational (who gains or loses access to opportunities and resources), and epistemic (whose knowledge is recognised, and who can contest a system’s account). On this understanding, gender bias is neither a mere defect to be engineered away nor an automatic verdict of structural inequality; it is a sociotechnical outcome whose specific form must be identified for a given system, task, and context.
The review proceeds as follows. Section 2 explains the construction of the literature corpus and the conceptual method. Section 3 reconstructs historical genealogies of gendered knowledge. Section 4 describes how GenAI systems generate outputs and where bias can enter the pipeline. Section 5 examines benchmarks and their limitations. Section 6 synthesizes evidence from employment, education, health, and multilingual or multimodal systems. Section 7 evaluates technical mitigation, auditing, and legal governance. Section 8 considers epistemic centralization, temporal orientation, and feminist alternatives. The final sections identify the paper’s conceptual contribution and research gaps.
This study is a critical integrative review and conceptual analysis. It is not a systematic review, meta-analysis, or scoping review, and it does not use Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) terminology. The corpus was assembled from the literature supplied by the authors together with targeted searches of Google Scholar, JSTOR, the Association for Computational Linguistics (ACL) Anthology, institutional repositories, and official policy sources; during the revision process, the search strategy was made explicit and the searches were extended and documented. The search was completed in July 2026. Sources in English and Italian were prioritized, with selected Spanish-language materials and research concerning other linguistic contexts included when they contributed directly to the analytical questions. Searches combined terms such as ‘gender bias’, ‘algorithmic bias’, ‘large language models’, ‘text-to-image’, ‘word embeddings’, ‘fairness’, ‘benchmark’, ‘feminist’, ‘intersectionality’, and ‘governance’, and were iterated as new terminology emerged. Scopus and Web of Science were not used, and the review does not claim exhaustive coverage of the literature. The final corpus comprised 66 sources.
The main temporal focus was 2013–2026, corresponding to the growth of embedding-based bias research, transformer models, and contemporary GenAI. Earlier works were retained when they supplied foundational concepts from feminist epistemology, Science and Technology Studies (STS), digital sociology, decolonial thought, and design justice [1][2][3][4][5][6][7][8][40][41][42][43][1,2,3,4,5,6,7,8,40,41,42,43]. Eligible materials included peer-reviewed empirical studies; benchmark and mitigation papers; systematic reviews and major surveys; books and theoretical articles with established relevance; and official governance documents. Preprints and theses were used sparingly, identified as such, and never treated as conclusive evidence [39][44][45][39,44,45].
Materials were excluded when they were duplicated, tangential to gender or GenAI, available only as secondary summaries when an original source could be identified, or unable to support the claim for which they had previously been cited. Particular attention was given to source-claim matching. Findings from other studies are attributed with formulations such as ‘the study reports’ or ‘the literature indicates’. Statements introduced as ‘this article argues’ identify the authors’ conceptual synthesis rather than an empirical result. Records were screened first by title and abstract for relevance to gender and generative AI, and then by full text for whether a source could substantiate the specific claim for which it was cited. Because this is a conceptual and integrative review rather than a systematic review, sources were appraised for relevance, evidential status, and the traceability of the claims drawn from them, rather than scored against a standardized risk-of-bias instrument designed for the aggregation of homogeneous studies.
The analysis used conceptual coding across five dimensions: (1) historical exclusions and data asymmetries; (2) technical representation and generation; (3) measurement and benchmark validity; (4) domain-specific deployment; and (5) mitigation, accountability, and epistemic governance. Reading continued until additional sources refined examples but no longer altered these five analytical categories. This is described as conceptual saturation, not exhaustive coverage of all publications. In this review, conceptual saturation denotes the stability of these five analytical categories under continued reading rather than completeness of the literature; this boundary is restated among the limitations.
Rigour is pursued through argumentative transparency, theoretical traceability, contestability, conceptual robustness, and empirical translatability. Argumentative transparency requires separating evidence, interpretation, and normative proposal. Theoretical traceability requires naming the traditions from which concepts such as situated knowledge, data colonialism, or design justice are drawn. Contestability requires acknowledging contradictory findings and the dependence of results on test design. Empirical translatability asks whether conceptual claims can guide measurable audits or institutional procedures without pretending that a conceptual framework has itself been experimentally validated. The construction of the corpus, analytical procedure, and rigour criteria are summarized in Table 1.
Table 1. Construction of the corpus and analytical procedure.
| Dimension | Specification |
|---|---|
| Review type | Critical integrative review and conceptual analysis; not a systematic review or meta-analysis. |
| Corpus construction | Author-supplied literature plus targeted searches in Google Scholar, JSTOR, ACL Anthology, institutional repositories, and official policy sources; search completed July 2026. |
| Coverage | Mainly 2013–2026, with earlier foundational works; English and Italian sources, with selected Spanish-language and multilingual materials. |
| Inclusion | Peer-reviewed empirical studies, benchmarks, major surveys, relevant theoretical works, and official governance documents. |
| Exclusion | Duplicates, tangential materials, unsupported summaries, and sources unable to support the specific claim. |
| Synthesis | Conceptual coding across genealogy, technical mechanisms, measurement, deployment, and governance; conceptual saturation rather than exhaustive enumeration. |
| Rigour criteria | Argumentative transparency, theoretical traceability, contestability, conceptual robustness, and empirical translatability. |