Introduction
Debates about gender and artificial intelligence (AI) usually begin with bias: hiring tools that favour men, image generators that picture nurses as women and engineers as men, facial analysis systems that fail darker-skinned women. These examples are real and important. But framing the problem mainly as "biased data" suggests a mainly technical solution: clean the data, adjust the model, and fairness will follow. This essay argues that technical interventions are necessary but insufficient. They can and do reduce discriminatory outcomes, yet the problems they address are themselves produced within institutional and economic structures. Gender inequality in AI is therefore best understood as a question of power: who designs these systems, whose lives become data, who performs the hidden labour and who captures the value, who decides where systems are deployed, and who can challenge them when they cause harm.
The essay develops this argument through a four-part framework of design, data, deployment and accountability power, with the distribution of economic value running through all four. It pays particular attention to Africa, where AI is arriving largely as an imported technology, shaped by priorities set elsewhere, and where African institutions are beginning to test whether they can hold global technology firms to account.
A framework for power in AI
The claim that AI inequality is "about power" needs a mechanism, not just a slogan. The causal chain can be stated in six steps. Existing social inequality shapes the data that is available and the assumptions designers bring to a problem. Those data and design choices shape how an AI system behaves. Someone then decides to deploy the system in a particular setting. Deployment distributes benefits and risks across different groups. Finally, whether harmed groups can contest the outcome depends on the accountability mechanisms available to them. At each step, a person or institution makes a choice, and gender inequality in outcomes can be traced to who holds that choice.
This approach draws on feminist scholarship on data and technology, which argues that analysing who holds power, and challenging how it is exercised, is the starting point for any claim to fairness (D'Ignazio and Klein, 2020). The framework used here distinguishes four forms of power. Design power determines what a system is built to do and whose experience counts as the default. Data power determines what becomes data, what is excluded and how it may be used. Deployment power determines where, on whom and for what purpose a system is used. Accountability power determines who can scrutinise, challenge and obtain remedy. The sections that follow treat each in turn.
Design power
How design choices encode gender
Machine learning systems learn from historical records, and those who design them decide which records to use and what counts as success. Amazon's experimental recruitment tool, abandoned in 2018, learned from a decade of male-dominated applications to penalise CVs containing the word "women's" (Dastin, 2018). The system did what it was designed to do: predict which applicants resembled past hires. The gendered outcome followed from a design decision to treat past hiring as the measure of merit.
Language technologies show similar patterns. Research on word embeddings found that models encoded analogies such as "man is to computer programmer as woman is to homemaker" (Bolukbasi et al., 2016), and a 2024 UNESCO study found that large language models associated women with domestic and caregiving roles far more often than men, while linking male names with careers, business and leadership (UNESCO, 2024). Because generative tools now draft job descriptions, lesson plans and policy text, their stereotypes spread quietly into institutional writing.
Facial technologies illustrate why precision about design matters. Buolamwini and Gebru's (2018) Gender Shades study tested gender classification, the task of labelling a face as male or female, and found error rates of up to 34.7 per cent for darker-skinned women compared with under 1 per cent for lighter-skinned men. A separate evaluation by the US National Institute of Standards and Technology examined face recognition, the task of matching one face to another, and found that many algorithms produced higher false match rates for women and for people of African and East Asian descent (Grother et al., 2019). These studies concern different tasks and their error rates are not directly comparable. What they share is a finding that systems designed and tested mainly on some populations perform worse for others.
Design choices can also exclude people entirely. Automated gender recognition systems typically assume two genders inferred from physical appearance. Keyes (2018) has shown that this design choice structurally erases trans and non-binary people and misgenders them wherever the technology is used for verification, access control or targeted content. The harm here is not an error in execution; it is written into the specification.
Representation without power
A common response is that more women in AI would produce fairer design. Women remain a minority of the field. A 2025 World Economic Forum and LinkedIn analysis found that although the gender gap in AI skills narrowed in almost every economy studied between 2018 and 2025, women still made up under a third of AI-skilled professionals on the platform, and held only about 12 per cent of STEM-related C-suite roles in 2024 (World Economic Forum, 2025). downtoearth
Diverse teams are more likely to notice harms that homogeneous teams overlook. But representation without decision-making authority has limits. In 2020, Timnit Gebru, co-lead of Google's Ethical AI team and co-author of Gender Shades, left the company in a dispute over a paper warning of the risks of large language models, including their tendency to encode social bias (Bender et al., 2021; Hao, 2020). Gebru has described her departure as a firing; Google characterised it as a resignation. The episode has been widely interpreted as illustrating the limits of individual researchers' influence within corporate AI governance. The stronger formulation of the representation argument is therefore that representation is necessary but insufficient unless it comes with decision-making authority, institutional accountability and control over data and deployment.
Fairness as a normative choice
Design power also includes the power to define fairness itself. Computer scientists have shown that several common mathematical definitions of fairness cannot all be satisfied at once, except in unusual circumstances (Kleinberg et al., 2016; Chouldechova, 2017). This technical result becomes a political one through four further steps. Different fairness criteria encode different values: one might prioritise equal error rates across groups, another equal accuracy of predictions. Choosing one criterion over another shifts errors, and therefore costs, from one group to another. Deciding which group should bear those costs is a normative judgement about justice. And normative judgements cannot be settled by technical optimisation alone. Whoever chooses the fairness definition is exercising power over the distribution of harm, which is why affected people, not developers alone, should take part in that choice.
Data power
Invisibility
Data power begins with who is present in the data at all. Women's lower digital participation has a direct mechanism of effect. According to GSMA's 2026 report, women in low- and middle-income countries are 12 per cent less likely than men to use mobile internet, and more than two-thirds of the 810 million women still unconnected live in South Asia and Sub-Saharan Africa. In Sub-Saharan Africa the gap narrowed from 30 per cent in 2024 to 26 per cent in 2025, a marked improvement on 38 per cent in 2018, yet some 230 million women in the region remained unconnected (GSMA, 2026). GSMAConnecting Africa
The consequence for AI follows in sequence. Lower digital participation produces fewer observations of women's behaviour and needs. Fewer observations mean weaker representation in the datasets used to train and evaluate models. Weaker representation can make automated decisions less accurate or less relevant for women, and it makes their circumstances less visible to the policymakers and firms relying on those systems. Invisibility in the data becomes invisibility in policy.
Extraction and exposure
The obvious remedy, collecting more data about women and marginalised groups, carries its own danger. Data can be used not only to include people but to surveil and punish them. Human Rights Watch documented 45 cases of arbitrary arrest involving 40 LGBT people in Egypt, Jordan, Lebanon and Tunisia, finding that in every arrest security forces searched people's phones, mostly by force or under threat, to collect or even fabricate evidence for prosecution (Human Rights Watch, 2023). These cases involve manual digital targeting rather than AI. They matter for this essay because they show what data about sexual orientation and gender identity can be used for in hostile legal environments. Automated profiling and content analysis would allow such targeting to operate at far greater scale, which makes it a clear risk that AI governance in the region must anticipate. Human Rights Watch
This is the central tension of data power. Invisibility excludes people from benefits; visibility without protection exposes them to harm. The goal is therefore not maximum visibility but meaningful control: women, LGBTQI+ people and marginalised communities deciding how data about them is collected, used and shared.
The labour behind the data
Data does not label itself. Investigative reporting has documented the outsourcing of data labelling and content moderation to lower-paid workers in the Global South. In Kenya, data labellers employed through an outsourcing firm were paid less than two dollars an hour to label violent and sexually abusive text so that a leading chatbot could learn to filter it (Perrigo, 2023). Kenyan content moderators working for a Meta subcontractor have since brought litigation alleging exploitative conditions and inadequate mental health support. In 2024, Meta lost its appeal against a ruling that it could be sued in Kenya, allowing a case brought by some 185 moderators from several African countries to proceed (Associated Press, 2024). ABC News
These cases reveal a value chain that runs from data collection through labelling and model development to deployment and profit. The costs, including low pay and exposure to traumatic content, fall disproportionately on workers in countries like Kenya. The value, in the form of intellectual property, cloud infrastructure and platform revenue, accrues mainly to firms headquartered elsewhere. The gender composition of this workforce and its gendered effects are poorly documented, which is itself a gap that research and regulation should address. What is clear is that a governance framework concerned with who benefits from AI cannot ignore where the labour happens and who is paid for it.
Deployment power
Algorithmic credit and the gender gap
Deployment power is the power to decide where a system is used and on whom. Digital credit in Kenya illustrates what is at stake. Mobile lending products assess applicants automatically using mobile money and phone data, promising access to credit without collateral, which could particularly benefit women who control fewer assets. The evidence so far is mixed. CGAP survey data found that 55 per cent of Kenyan digital borrowers were men, and providers reported that men borrow more often and in larger amounts (Gubbins and Totolo, 2018). A nationally representative analysis found persistently lower shares of women in the formal digital credit market, and a gender gap in total formal credit use that grew from 4.1 percentage points in 2012, before digital credit launched, to 7.8 percentage points after its introduction (Johnen et al., 2023). CGAPResearchGate
These findings do not show that credit algorithms discriminate against women directly; lower uptake may reflect differences in phone ownership, mobile money use, income and preferences. The gender gap in mobile money account ownership in Sub-Saharan Africa stood at 25 per cent in 2024 (GSMA, 2025), and models trained on mobile money data inherit that gap. The point is that deploying automated credit scoring in a setting with deep existing inequalities, without monitoring gendered outcomes, can widen the very gap it promised to close. Scoring models that rely on phone usage, transaction histories or airtime purchases may disadvantage women who share handsets or earn informally. Whether they do so in practice is an empirical question that lenders and regulators are well placed to answer, and should be required to. GSMA
State deployment
Deployment decisions are made by states as well as firms. Governments increasingly use automated and biometric systems for identity, welfare eligibility, policing and border control. Kenya's attempt to roll out a centralised biometric identity system, known as Huduma Namba, shows both the risks and a route to accountability. In October 2021 the High Court quashed the government's decision to issue Huduma cards, ruling that the state had breached the Data Protection Act by failing to conduct a data protection impact assessment, and ordered it to carry one out before processing data (Privacy International, 2021). The case was not about gender specifically. It matters because such systems determine access to services, and women, who are often less likely to hold identity documents, bear particular risks of exclusion when registration becomes a condition of access. hapakenya
Labour market effects
Deployment also reshapes work. The World Economic Forum and LinkedIn found that women were more likely to hold roles disrupted by generative AI and less likely to benefit from AI-driven career augmentation (World Economic Forum, 2025). Decisions about which tasks to automate, and whether affected workers are retrained, are deployment decisions with gendered consequences. downtoearth
Accountability power
Technology-facilitated violence
The weakest link in the chain is often accountability: the ability of harmed people to obtain redress. AI has created new tools for gender-based violence. An early analysis found that non-consensual pornography accounted for 96 per cent of the deepfake videos it identified online, and that the subjects of the pornographic deepfakes on the leading dedicated websites it examined were all women (Ajder et al., 2019). An industry report by a commercial security website later described a sharp rise in deepfake videos between 2019 and 2023, with pornography and women targets still predominant (Home Security Heroes, 2023); as grey literature, it indicates a trend rather than establishing a precise figure.
African evidence shows how weak accountability compounds these harms. Pollicy's study, drawing on 3,306 surveys, interviews and focus groups with women in Ethiopia, Kenya, Uganda, Senegal and South Africa, found that 28 per cent of women had experienced some form of online harassment, and 40 per cent of those believed gender was a primary reason. Most incidents occurred on Facebook, yet up to 95 per cent of the women were unaware of any laws or policies protecting them against online gender-based violence (Iyer et al., 2020). Harm without known remedy is a direct expression of accountability deficits. An African Feminist Perspective: Online gender-based violence +2
Women in public life face particular exposure. A global survey found that nearly three quarters of women journalists had experienced online violence in their work (Posetti et al., 2021). This essay projects, as a risk rather than an established trend, that synthetic sexual imagery will increasingly feature in such campaigns. Technology-facilitated abuse also operates in private: research with survivors of intimate partner violence has documented how abusers exploit ordinary apps, shared accounts and location tracking (Freed et al., 2018), and AI-driven analysis of personal data could make such monitoring more powerful. For survivors facing weak cyber laws, platforms slow to act on reports in local languages, and heavy social stigma, redress can be almost unattainable.
African institutions asserting accountability
Accountability is not only absent. The Kenyan cases discussed above show African courts asserting jurisdiction over global technology firms and requiring the state to assess data protection risks before deploying biometric systems. These precedents matter for gender because accountability mechanisms, once established, can be used by women and LGBTQI+ people to challenge harmful systems. They also demonstrate that African institutions need not be passive recipients of technology governed elsewhere.
The possibilities, on different terms
None of this means AI is inherently harmful to gender equality. The question is who holds power over it. Masakhane, a grassroots network of African researchers, has built natural language processing tools for African languages through participatory, open methods that treat speakers as collaborators rather than data sources (Nekoto et al., 2020). It represents a shift from Africa as a site of data extraction to Africa as a producer of AI. Such tools matter for gender equality because in countries where women and girls face gaps in literacy and digital access, voice-based services in local languages could deliver agricultural, health and legal information to women excluded by text-heavy, English-language platforms.
AI can also analyse pay and recruitment data to reveal discrimination, process large volumes of community feedback to strengthen services, and provide confidential first contact points for survivors who cannot safely approach services in person. In each case the same conditions apply: tools designed with affected communities, built on survivor-centred and do-no-harm principles, and linked to real human services and referral pathways rather than replacing them.
Governance: four pillars
If gender inequality in AI is a problem of power, governance is the means of redistributing it. International frameworks provide a foundation. The UNESCO Recommendation on the Ethics of Artificial Intelligence (UNESCO, 2021), adopted by all member states, contains a dedicated policy area on gender. The European Union's AI Act (European Union, 2024) lists in Annex III several categories of high-risk use subject to risk management, documentation and human oversight requirements, including systems used in employment and worker management, systems that evaluate the creditworthiness of natural persons or establish their credit score, and systems that determine access to essential public and private services and benefits. The African Union's Continental Artificial Intelligence Strategy (African Union, 2024) calls for AI development that is inclusive, responsible and rooted in African development priorities. Turning these commitments into practice requires action across four pillars, each corresponding to a form of power identified above.
Assessment. Governments should require gender and intersectional impact assessments for high-risk AI systems before deployment, with algorithmic testing disaggregated by sex, age, disability and other relevant characteristics, and repeated after deployment to detect drift. Results should be published. Lenders using automated credit scoring, for example, should be required to report outcomes by gender.
Accountability. Systems should be auditable by independent bodies, and affected people should have accessible complaint procedures and effective remedies. Non-consensual intimate imagery, including synthetic imagery, requires specific legal prohibition and rapid takedown obligations on platforms, with reporting channels that function in local languages. Consequential automated decisions should remain subject to meaningful human oversight.
Participation and power. Women's rights organisations, LGBTQI+ organisations, affected communities and African researchers should hold formal roles in drafting national AI strategies, setting fairness criteria for high-risk systems and overseeing public procurement, not merely consultative status after decisions are made.
Inclusion and capability. States and development partners should invest in affordable connectivity and digital skills for women and girls, in local AI research and development, and in the regulatory and procurement capacity needed to scrutinise systems supplied by foreign firms. Data protection authorities need the resources to enforce the law against foreign as well as domestic actors.
Conclusion
The central question about gender and AI is not whether the technology is biased. It frequently is, and technical work to measure and reduce that bias remains essential. The more important question is who holds power at each stage: who designs systems and defines fairness, whose lives become data and on what terms, who performs the labour and captures the value, who decides where systems are deployed, and who can challenge them when they cause harm. Technical debiasing addresses symptoms within that chain; governance, participation and accountability address its causes.
For Africa, the timing matters. Many countries are now drafting AI strategies, data protection regimes and digital public infrastructure, and the rules set in this period will shape AI for decades. Kenyan courts have already shown that global firms and national governments can be held to account. Gender equality advocates have often entered technology debates late, responding to harms after systems are entrenched. The present moment offers a chance to act earlier: to write gender analysis, budgets for inclusion and the participation of women and gender-diverse people into these frameworks from the start. If they remain positioned mainly as data sources and end users, AI risks reproducing and amplifying the inequalities it learns from. If they become co-designers, auditors and rule-makers, AI can become a tool for challenging those inequalities rather than automating them.
References
African Union (2024). Continental Artificial Intelligence Strategy. Addis Ababa: African Union.
Ajder, H., Patrini, G., Cavalli, F. and Cullen, L. (2019). The State of Deepfakes: Landscape, Threats, and Impact. Amsterdam: Deeptrace.
Associated Press (2024). Facebook loses jurisdiction appeal in Kenyan court paving the way for moderators' case to proceed. AP News, September.
Bender, E.M., Gebru, T., McMillan-Major, A. and Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pp. 610 to 623.
Bolukbasi, T., Chang, K.W., Zou, J., Saligrama, V. and Kalai, A. (2016). Man is to computer programmer as woman is to homemaker? Debiasing word embeddings. Advances in Neural Information Processing Systems, 29.
Buolamwini, J. and Gebru, T. (2018). Gender shades: Intersectional accuracy disparities in commercial gender classification. Proceedings of Machine Learning Research, 81, pp. 77 to 91.
Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5(2), pp. 153 to 163.
Dastin, J. (2018). Amazon scraps secret AI recruiting tool that showed bias against women. Reuters, 10 October.
D'Ignazio, C. and Klein, L.F. (2020). Data Feminism. Cambridge, MA: MIT Press.
European Union (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act), Annex III. Official Journal of the European Union.
Freed, D., Palmer, J., Minchala, D., Levy, K., Ristenpart, T. and Dell, N. (2018). "A stalker's paradise": How intimate partner abusers exploit technology. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, paper 667.
Grother, P., Ngan, M. and Hanaoka, K. (2019). Face Recognition Vendor Test (FRVT) Part 3: Demographic Effects. NISTIR 8280. Gaithersburg, MD: National Institute of Standards and Technology.
GSMA (2025). Progress in closing the mobile money gender gap has stalled: Latest evidence from Findex 2025. Mobile for Development blog, November. London: GSMA.
GSMA (2026). The Mobile Gender Gap Report 2026. London: GSMA.
Gubbins, P. and Totolo, E. (2018). Kenya's digital credit revolution five years on. CGAP blog, 15 March. Washington, DC: CGAP.
Hao, K. (2020). We read the paper that forced Timnit Gebru out of Google. Here's what it says. MIT Technology Review, 4 December.
Home Security Heroes (2023). 2023 State of Deepfakes: Realities, Threats, and Impact. Industry report.
Human Rights Watch (2023). "All This Terror Because of a Photo": Digital Targeting and Its Offline Consequences for LGBT People in the Middle East and North Africa. New York: Human Rights Watch.
Iyer, N., Nyamwire, B. and Nabulega, S. (2020). Alternate Realities, Alternate Internets: African Feminist Research for a Feminist Internet. Kampala: Pollicy.
Johnen, C., Parlasca, M. and Mußhoff, O. (2023). Digital credit and the gender gap in financial inclusion: Empirical evidence from Kenya. Journal of International Development. doi:10.1002/jid.3687.
Keyes, O. (2018). The misgendering machines: Trans/HCI implications of automatic gender recognition. Proceedings of the ACM on Human-Computer Interaction, 2(CSCW), article 88.
Kleinberg, J., Mullainathan, S. and Raghavan, M. (2016). Inherent trade-offs in the fair determination of risk scores. arXiv preprint arXiv:1609.05807.
Nekoto, W. et al. (2020). Participatory research for low-resourced machine translation: A case study in African languages. In Findings of the Association for Computational Linguistics: EMNLP 2020, pp. 2144 to 2160.
Perrigo, B. (2023). OpenAI used Kenyan workers on less than $2 per hour to make ChatGPT less toxic. TIME, 18 January.
Posetti, J., Shabbir, N., Maynard, D., Bontcheva, K. and Aboulez, N. (2021). The Chilling: Global Trends in Online Violence against Women Journalists. Paris: UNESCO.
Privacy International (2021). Data protection impact assessments and ID systems: The 2021 Kenyan ruling on Huduma Namba. London: Privacy International.
UNESCO (2021). Recommendation on the Ethics of Artificial Intelligence. Paris: UNESCO.
UNESCO (2024). Challenging Systematic Prejudices: An Investigation into Bias Against Women and Girls in Large Language Models. Paris: UNESCO.
World Economic Forum (2025). Gender Parity in the Intelligent Age. Geneva: World Economic Forum, with LinkedIn.
