>>AI & Gender
The drive for automation increasingly shapes every aspect of contemporary life, with technological artefacts and infrastructures structuring our actions, behaviours, and interactions. Artificial Intelligence (AI), in particular, has become a central technology; AI is involved deeply in everything from healthcare access to hiring decisions, criminal sentencing, and everyday communication. Rather than viewing technology through a deterministic lens - as neutral, autonomous, and detached from social interests - recent scholarship highlights a model of co-shaping: society shapes technology while technology, simultaneously, reshapes society.
Gender stands in a mutually constitutive relationship with technology. Engendered ideologies become embedded in the design, production and use of technologies; this, in turn, reinforces cultural norms, values, and expectations about gender. Indicatively, feminist scholars caution against what D’Ignazio and Klein (2020) conceptualise as “Big Dick Data” —a term describing grand data-driven projects built on masculine fantasies of objectivity, scale, and control, while ignoring context, culture, and lived experience.
Gender biases in AI technologies typically emerge at three interconnected levels:
Design and Programming: Biases introduced through interface-level or implementation-level human decisions. These include choices that shape how a system appears or behaves at the surface: for instance, selecting a default female voice for digital assistants, hard-coding binary gender categories into a classifier, or defining user workflows that implicitly assume a gendered norm. These are conscious, user-facing design decisions.
Training Data: Biases reproduced from the datasets used to train models. For example, speech-recognition systems trained predominantly on male voice recordings, image datasets that largely depict men in technical roles, or translation corpora that reflect gender stereotypes. Here, the model mirrors imbalances already present in the historical data.
Algorithmic Logic: Biases arising from modelling decisions at a deeper computational level. These relate to the mathematical concepts of the system—such as choices of optimisation algorithms, neural network architectures, clustering methods, or error-weighting schemes. Even when the dataset remains unchanged, such modelling choices can systematically amplify gender disparities. Example: an acoustic model in a speech-recognition system may optimise accuracy around the largest statistical cluster of voices (typically male), producing consistently higher error rates for women. These outcomes do not imply autonomous technological agency; rather, they reflect human decisions operating at an abstract mathematical layer that is less visible than interface design.
AI biases are not harmless glitches; they are systemic (Broussard, 2024). If a gender-related prejudice is encoded into the technology, it will be reproduced indefinitely. Empirical cases underline the issue. AI biases manifest through different but interconnected mechanisms. Some operate at the level of representation and design, while others emerge through data-driven and algorithmic processes.
At the representational level, voice assistants such as Siri, Alexa, Cortana, and Tao are overwhelmingly feminized—both in name and voice—reflecting a long history of feminized emotional labour, from 1950s telephone operators to contemporary service work. These design choices reproduce gendered expectations about care, compliance, and emotional availability. Yet, despite their “female” outlooks, such systems often fail to respond adequately to issues disproportionately affecting women, such as domestic violence.
At the data and algorithmic level, bias is embedded through historical datasets and modelling processes. Amazon’s 2018 hiring algorithm, for instance, systematically favoured male candidates because it was trained on résumés reflecting a decade of male-dominated hiring practices in the tech industry.
Beyond these technical and representational layers, gender bias in AI is also structurally reinforced through the male domination of AI-related labour and expertise. The overrepresentation of men in AI research, engineering, and decision-making roles shapes what problems are prioritised, which datasets are collected, and how “success” is defined. This institutional imbalance operates upstream of both design and algorithmic bias, influencing AI systems long before deployment.
Emojis, governed by the Unicode Consortium, also reveal gender bias (Pérez, 2019). While emojis themselves are not AI systems, they constitute training and representational infrastructures for AI applications, including sentiment analysis, natural language processing, and text-to-image models. Gendered emoji representations, therefore, contribute to the reproduction of bias within AI-mediated communication systems.
Addressing gender bias in AI requires a shift from reactive correction to proactive redesign. This involves:
- Inclusive design teams that reflect diverse social identities and experiences. Inclusive design teams that reflect diverse social identities and experiences, not as a symbolic fix to male domination, but as a structural intervention that broadens problem-framing, data selection, and design priorities throughout the AI development process.
- Algorithmic/dataset transparency & gender bias auditing to identify imbalances before deployment. Algorithmic and dataset transparency, alongside systematic gender bias auditing, to mitigate the inheritance and amplification of gender biases embedded in training data and modelling choices before deployment.
- Context-aware development, integrating cultural and gender-sensitive analysis throughout the entire AI lifecycle.
- Regulatory and ethical frameworks ensuring accountability for gender-related discriminatory outcomes.
We maintain that overcoming gender bias in AI demands a shift from post-hoc correction to proactive responsibility. This entails inclusive and accountable design practices, critical engagement with training data and algorithmic choices, and institutional reforms that address the gendered organisation of AI labour. Without such systemic interventions, AI systems risk perpetuating and amplifying existing inequalities.
REFERENCES
Broussard, M. (2024). More than a Glitch: Confronting Race, Gender, and Ability Bias in Tech. MIT Press.
Broussard, M. (2018). Artificial unintelligence: How computers misunderstand the world. MIT Press.
Browne, J., Cave, S., Drage, E., & McInerney, K. (2023). Feminist AI: Critical Perspectives on Algorithms, Data, and Intelligent Machines. Oxford University Press.
Manasi, A., Panchanadeswaran, S., Sours, E., & Lee, S. J. (2023). Mirroring the bias: gender and artificial intelligence. Gender, Technology and Development, 26(3), 295-305.
Pérez, C.C. (2019). Invisible Women: Data Bias in a World Designed for Men. Abrams Press.