Whose Voice Does AI Hear? Gender, Language and Representation in Generative AI

Whose Voice Does AI Hear? Gender, Language and Representation in Generative AI

Introduction

Artificial Intelligence (AI) has become an important part of contemporary communication, education, research, creative writing, and digital culture. The emergence of generative AI has introduced new possibilities for producing essays, stories, poems, dialogues, summaries, and other forms of textual communication. Generative AI systems are capable of producing human-like language in response to user prompts and are increasingly being used in academic, professional, and creative contexts.

Large Language Models (LLMs) generate language by learning patterns from extensive collections of human-produced texts. As language is closely connected with culture, society, identity, and power, AI-generated language can reflect some of the assumptions and patterns present in the data on which these systems are trained. Consequently, the study of AI-generated language is not only a technological concern but also a linguistic, cultural, and social concern.

Gender is one of the important areas in which language contributes to the construction of identity. Literary and cultural texts have historically represented women through particular social roles and expectations. Women have frequently been associated with domesticity, motherhood, emotionality, beauty, care, and dependence, while men have often been represented through authority, rationality, leadership, independence, and professional success. Contemporary literature has challenged many of these conventional representations, yet gendered patterns continue to exist in different forms of discourse.

The emergence of generative AI raises an important question: When AI generates a story about a woman, whose understanding of womanhood is being represented? The question becomes particularly significant because AI-generated narratives can appear neutral, objective, and authoritative even when they contain culturally inherited assumptions.

Language plays a central role in the construction of gender. The choice of adjectives, verbs, nouns, pronouns, metaphors, descriptions, speech patterns, and grammatical structures can influence how readers perceive characters and identities. For example, describing a female character as “beautiful,” “caring,” or “emotional” creates a different representation from describing her as “decisive,” “strategic,” “assertive,” or “authoritative.” Similarly, the grammatical representation of women as actors or recipients of actions can influence the construction of agency.

Generative AI therefore provides a new field for linguistic investigation. Instead of examining only human-authored literary texts, researchers can now investigate texts produced by machines and ask how gender is linguistically constructed within them.

Background Study

Gender and language have been extensively studied within sociolinguistics, discourse studies, feminist linguistics, and literary studies. Language does not merely communicate information; it can also construct identities and social relationships.

Gendered language can occur through vocabulary, grammatical structures, forms of address, conversational practices, metaphors, descriptions, and patterns of representation. Literary texts, advertisements, media discourse, political discourse, and everyday communication may reproduce or challenge gender expectations through linguistic choices.

Feminist linguistic approaches have particularly examined how language can contribute to the marginalisation of women and how women can reclaim linguistic agency. The study of women's voices therefore involves not only examining whether women are present in a text but also examining how they are allowed to speak, what they are represented as saying, and how much authority their voices possess.

Generative AI refers to artificial intelligence systems capable of producing new content based on patterns learned from existing data. Large Language Models can generate human-like textual responses to prompts and can be used to create narratives, dialogues, essays, poems, and other forms of discourse.

The increasing use of generative AI raises questions about authorship, originality, representation, bias, and linguistic agency. Although AI does not possess human experiences in the conventional sense, its outputs may reproduce linguistic patterns associated with human societies.

This makes AI-generated text an important object of linguistic study. The question is no longer simply whether AI can produce grammatically correct language, but also what kinds of social meanings are produced through that language.

Gender representation refers to the ways in which individuals of different genders are portrayed through language, images, narratives, and cultural discourse.

In AI-generated narratives, gender representation can be examined through:

• character descriptions;

• occupations and social roles;

• personality traits;

• physical appearance;

• emotional characteristics;

• dialogue;

• decision-making;

• leadership;

• relationships;

• domestic responsibilities;

• professional identities; and

• representations of power and authority.

A central concern of the present study is whether generative AI reproduces conventional gender stereotypes or creates more diverse representations of women.

Women’s Voice and Agency

Voice refers to the ability of an individual to express thoughts, opinions, experiences, and identities. In literary discourse, voice is closely associated with agency and power.

A female character may appear prominently in a narrative but may not necessarily possess narrative agency. For example, she may be described extensively by other characters while having limited opportunities to speak for herself.

The present study therefore examines not only how often women speak, but also:

• What do women say?

• Who listens to them?

• Are their opinions respected?

• Are their decisions represented as important?

• Are they portrayed as active decision-makers?

• How are their emotions described?

• Are their voices independent or mediated through male characters?

Gender Stereotypes

Gender stereotypes are socially constructed assumptions about the characteristics, behaviours, roles, and abilities associated with different genders.

AI-generated narratives may reproduce stereotypes because language models learn from large quantities of existing human language. At the same time, carefully designed prompts may generate narratives that challenge traditional gender roles.

The study therefore investigates both possibilities rather than assuming in advance that AI is either inherently biased or completely neutral.

Objectives

i. To examine the linguistic representation of women in selected AI-generated narratives.

ii. To analyse the construction of female voice and agency in generative AI discourse.

iii. To identify gendered lexical and grammatical patterns in AI-generated narratives.

iv. To examine the representation of women in relation to power, identity, profession, relationships, and social expectations.

Methodology

The study adopts a descriptive and qualitative research approach supported by linguistic and discourse analysis. Selected narratives will be generated using generative AI tools through carefully designed prompts focusing on women's identities, relationships, professional roles, leadership, social expectations, and resistance.

The generated texts will constitute the primary corpus for analysis. The study may compare responses produced by two or more generative AI platforms using similar prompts.

The analysis will employ principles from Critical Discourse Analysis, Feminist Linguistics, Systemic Functional Linguistics, and Corpus Linguistics.

Particular attention will be given to:

• transitivity and agency;

• lexical choices;

• adjectives and evaluative language;

• modality;

• pronoun usage;

• direct and indirect speech;

• representation of emotions;

• active and passive constructions;

• descriptions of physical appearance;

• professional and domestic roles; and

• expressions of authority and resistance.

The study will compare patterns across the selected AI-generated texts to identify recurring linguistic representations.

Conceptual Framework

The study can be situated within three major theoretical perspectives:

Feminist Linguistics

Feminist Linguistics examines the relationship between language, gender, power, and social structures. It provides a framework for examining how linguistic choices may reproduce or challenge gender inequality.

Critical Discourse Analysis

Critical Discourse Analysis examines how language is connected with power, ideology, social structures, and representation. It is particularly useful for investigating how AI-generated discourse constructs gender identities.

Systemic Functional Linguistics

Systemic Functional Linguistics provides tools for examining how language creates meanings through processes, participants, circumstances, modality, and interpersonal relationships. The transitivity system can be particularly useful for analysing whether female characters are represented as active participants or passive recipients.

Literature Review

The literature review may be organised under the following areas:

1. Gender and Language

2. Feminist Linguistics

3. Women and Literary Representation

4. Gender and Critical Discourse Analysis

5. Artificial Intelligence and Language

6. Gender Bias in Artificial Intelligence

7. Large Language Models and Linguistic Representation

8. Generative AI and Creative Writing

9. AI-Generated Narratives and Gender

10. AI Ethics, Representation and Algorithmic Bias

Gender and Language

• Academic Framing: Focus on the evolution of sociolinguistic frameworks (e.g., the deficit, dominance, difference, and dynamic approaches).

• Literature Review Example: Early sociolinguistic research often framed women’s language through the lens of deficit or difference. For instance, classic empirical studies established that women frequently employ tag questions ("isn't it?") and hedging devices ("perhaps") to foster conversational collaboration. Conversely, men were found to dominate turn-taking and utilize competitive interruption patterns to assert institutional hierarchy. Modern scholars, however, argue that these features are context-dependent rather than inherently gendered.

2. Feminist Linguistics

• Academic Framing: Critique the structural sexism embedded within the morphology and syntax of language itself.

• Literature Review Example: Grounded in patriarchal critique, feminist linguists demonstrate how natural language operates as a gendered hierarchy. Scholars like Robin Lakoff exposed how structural grammar positions the masculine as the semantic default. This is evident in the historic institutionalization of generic masculine pronouns ("he" to mean anyone) and androcentric lexical items ("mankind", "chairman"). This structural erasure effectively marginalizes non-male identities from public and legal domains.

3. Women and Literary Representation

• Academic Framing: Trace how cultural narratives construct and police gender roles through written texts.

• Literature Review Example: Literary criticism has long documented the polarization of female characters into rigid archetypes—most notably the "angel in the house" or the "madwoman in the attic." An examination of 19th-century canonical fiction reveals that female character arcs are routinely restricted to domestic, emotional, or passive spheres. Meanwhile, agency, intellect, and external mobility are systematically reserved for their male counterparts. This established a literary precedent for how gendered behavior is culturally normalized.

4. Gender and Critical Discourse Analysis (CDA)

• Academic Framing: Show how everyday language and media reinforce institutional power structures.

• Literature Review Example: Applying Feminist Critical Discourse Analysis (FPCDA) exposes how media and political texts subtly legitimize gender inequalities. CDA studies of corporate press releases show a persistent linguistic asymmetry: male executives are consistently modified with agentic, authoritative verbs ("spearheaded", "commanded"), whereas female executives are evaluated using soft skills or communal descriptors ("nurtured", "collaborated"), reinforcing glass ceilings through discourse.

5. Artificial Intelligence and Language

• Academic Framing: Explain how Natural Language Processing (NLP) inherits human social patterns through data training.

• Literature Review Example: The transition of human language into machine-readable data has carried historical biases into computer science. Because NLP models rely on distribution-based semantic vector spaces (word embeddings), they map words based on human cultural proximity. Consequently, foundational computational linguistics research reveals that algorithms mathematically link the concept of "man" to "computer programmer" and "woman" to "homemaker," proving that AI mirrors human cognitive biases.

6. Gender Bias in Artificial Intelligence

• Academic Framing: Detail the real-world, discriminatory consequences of automated decision-making.

• Literature Review Example: Machine learning algorithms frequently amplify systemic biases hidden within training datasets. In corporate settings, automated resume-screening algorithms trained on historical tech-industry hiring data learned to systematically penalize applications containing the word "women's" (e.g., "women's chess club captain"). This demonstrates how predictive AI can actively replicate and scale historical workplace discrimination.

7. Large Language Models (LLMs) and Linguistic Representation

• Academic Framing: Analyze the statistical probabilities that dictate how modern generative models output gendered text.

• Literature Review Example: Despite advanced optimization, Large Language Models (LLMs) output text based on the statistical probabilities of their training corpora. Literature evaluating models like GPT or LLaMA reveals that when generating blind professional profiles, the models exhibit statistical skewing. They consistently default to masculine pronouns for highly compensated STEM fields and feminine pronouns for pink-collar occupations, demonstrating a regression to stereotypical linguistic norms.

8. Generative AI and Creative Writing

• Academic Framing: Examine human-AI co-creativity and the stylistic constraints of machine text.

• Literature Review Example: The integration of generative AI into creative writing introduces a tension between human imagination and algorithmic predictability. Literary studies investigating co-authored human-AI fiction find that generative tools tend to smooth out subversive syntax. Instead, they favor safe, highly frequent generic tropes. The AI's stylistic suggestions often default to conventional narrative structures, limiting the stylistic experimentation vital to progressive creative writing.

9. AI-Generated Narratives and Gender

• Academic Framing: Critically analyze the plotlines, characters, and tropes written entirely by AI.

• Literature Review Example: Narratology in the age of AI reveals that fully machine-generated stories are highly susceptible to narrative stereotyping. When prompted to generate fiction without explicit demographic constraints, AI models routinely generate traditional nuclear family scripts. In these stories, male characters are assigned external, problem-solving, and adventurous roles, while female characters are relegated to internal emotional management and caretaking, thereby mass-producing outdated gender roles.

10. AI Ethics, Representation, and Algorithmic Bias

• Academic Framing: Propose the theoretical frameworks, policy revisions, and active interventions needed for equitable AI.

• Literature Review Example: Current scholarship in AI ethics argues that technological neutrality is a myth. Addressing digital patriarchy requires shifting from passive "data cleaning" to active intersectional feminist audits. Ethicists advocate for participatory design frameworks that involve marginalized communities in data curation, establishing strict regulatory guardrails to ensure LLMs preserve equitable linguistic representation.

Although considerable research has examined gender representation in literature, media, advertising, and digital communication, the rapid development of generative AI has created a new area for investigation. Existing studies on AI bias have frequently focused on technical performance, datasets, occupations, or general stereotypes. Comparatively less attention can be given to the linguistic construction of women's voices, agency, identity, and representation within AI-generated literary narratives. The present study therefore seeks to bridge the fields of gender studies, linguistics, literary discourse, and generative AI by examining how AI constructs women through language.

Data Analysis

The present study analyses selected AI-generated narratives to examine how women are represented through language, particularly in relation to voice, agency, identity, power, emotion, profession, relationships, and social expectations. The analysis combines qualitative discourse analysis with selected quantitative measures in order to identify recurring linguistic patterns. The data are examined through the theoretical perspectives of Feminist Linguistics, Critical Discourse Analysis, and Systemic Functional Linguistics.

The analysis is organised into the following major categories: (1) lexical representation, (2) transitivity and agency, (3) voice and speech representation, (4) evaluative and emotional language, (5) professional and domestic identities, (6) authority and decision-making, and (7) stereotypical and non-stereotypical representation.

Corpus of AI-Generated Narratives

The primary data consist of narratives generated through carefully designed prompts focusing on women's identities and experiences. The prompts address areas such as professional life, leadership, family relationships, social expectations, personal independence, conflict, and resistance.

To ensure comparability, the same or closely equivalent prompts may be submitted to different generative AI systems. The responses are collected without substantial modification so that the linguistic characteristics of the generated texts can be examined.

Each narrative is assigned a code, such as N1, N2, N3, and so forth. Where more than one AI platform is used, platform-based codes such as AI-A, AI-B, and AI-C may be adopted. This coding system allows the analysis to refer to individual texts without repeatedly naming the platforms.

Conclusion

The study examines the relationship between gender, language, and representation in generative AI. By analysing AI-generated narratives through feminist linguistic and discourse-based approaches, the study demonstrates how linguistic choices can contribute to the construction of female identity, agency, power, and social roles.

The analysis can reveal whether women are predominantly represented through conventional characteristics such as beauty, emotion, domesticity, and relationships or whether AI-generated narratives provide more diverse representations involving leadership, professional achievement, independence, resistance, and decision-making.

The study also highlights the importance of examining AI-generated language critically. Generative AI should not be treated as a completely neutral producer of language because its outputs may reflect patterns found in human-generated discourse. At the same time, AI-generated texts can potentially create alternative representations when prompts encourage diverse and non-stereotypical perspectives.

The study therefore establishes the importance of critical AI literacy and linguistic awareness in understanding machine-generated discourse. Examining whose voices are represented, how those voices are constructed, and which identities receive agency can contribute to more responsible and inclusive uses of generative AI.

Outcomes

1. The study identifies linguistic patterns associated with the representation of women in AI-generated narratives.

2. It examines the relationship between gender, voice, agency, and power in generative AI discourse.

3. It identifies instances of gender stereotyping and non-stereotypical representation.

4. It demonstrates the usefulness of feminist linguistics and discourse analysis in studying AI-generated texts.

5. It contributes to the emerging field of AI-mediated literary and linguistic studies.

6. It highlights the importance of critical AI literacy in recognising potential gender bias in machine-generated language.

Limitations

The study is limited to selected AI-generated narratives and selected generative AI platforms. AI systems are continuously updated, and therefore their outputs may change over time. The findings will also depend on the prompts used, the number of texts analysed, and the linguistic framework adopted.

The study does not attempt to determine the technical architecture or internal functioning of AI systems. Instead, it focuses primarily on the linguistic and discursive characteristics of their generated texts.

Future research may employ larger corpora, computational corpus analysis, multiple languages, and a wider range of AI systems.

References

• Bucholtz, M. (2014). The feminist foundations of language, gender, and sexuality research. In S. Ehrlich, M. Meyerhoff, & J. Holmes (Eds.), The handbook of language, gender, and sexuality (2nd ed.). Wiley.

• Halliday, M. A. K., & Matthiessen, C. M. I. M. (2014). Halliday's introduction to functional grammar (4th ed.). Routledge.

• Lakoff, R. (2008). Language and woman's place. Language in Society. Cambridge University Press.

• Lakoff, R. (2003). Language, gender, and politics: Putting “women” and “power” in the same sentence. In J. Holmes & M. Meyerhoff (Eds.), The handbook of language and gender. Wiley.

• Kotek, H., Dockum, R., & Sun, D. Q. (2023). Gender bias and stereotypes in large language models. Proceedings/Research paper.

• Valderrey Nuñez, A., & Bloem, J. (2026). Investigating how LLMs propagate female stereotypes: Comparing what models say via prompts with what they represent in their embeddings. Proceedings of LREC 2026.

• Rowe, J., Klimaszewski, M., Guillou, L., Vallor, S., & Birch, A. (2025). EuroGEST: Investigating gender stereotypes in multilingual language models. Proceedings of EMNLP 2025.

• UNESCO. (2024). Generative AI: UNESCO study reveals alarming evidence of regressive gender stereotypes. UNESCO.

• Bartley, L. V. (2018). Putting transitivity to the test: A review of the Sydney and Cardiff models. Functional Linguistics, 5, Article 4.

• Martin, J. R., & White, P. R. R. (2005). The language of evaluation: Appraisal in English. Palgrave Macmillan.

• Halliday, M. A. K., & Matthiessen, C. M. I. M. (1999). Construing experience through meaning: A language-based approach to cognition. Cassell.

• Matthiessen, C. M. I. M. (1999). The system of transitivity: An exploratory study of text-based profiles. Functions of Language, 6(1), 1–51.

• Holmes, J., & Meyerhoff, M. (Eds.). (2003). The handbook of language and gender. Wiley.

• Butler, J. (1990). Gender trouble: Feminism and the subversion of identity. Routledge.

• Fairclough, N. (1992). Discourse and social change. Polity Press.

• Fairclough, N. (1995). Critical discourse analysis: The critical study of language. Longman.

• van Dijk, T. A. (1993). Principles of critical discourse analysis. Discourse & Society, 4(2), 249–283.

• Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–623.

• Bolukbasi, T., Chang, K.-W., Zou, J., Saligrama, V., & Kalai, A. T. (2016). Man is to computer programmer as woman is to homemaker? Debiasing word embeddings. Advances in Neural Information Processing Systems, 29.

• Blodgett, S. L., Barocas, S., Daumé III, H., & Wallach, H. (2020). Language (technology) is power: A critical survey of “bias” in NLP. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5454–5476.