The AI System Completing Your Sentences Is Completing Them With a Gender Bias. Here Is the Scale of the Problem.
Large language models — the AI systems that millions of people now use to draft emails, plan campaigns, write content, and search for information — have been trained on human-generated text accumulated over decades. That text reflects the world as it was and, in many respects, as it still is: a world in which women are systematically underrepresented in positions of professional authority, overrepresented in domestic and caregiving contexts, and depicted through a set of gender stereotypes so pervasive that they have become structurally invisible to most people who encounter them.
When AI systems learn from that text, they learn the stereotypes along with the vocabulary. And when they generate new content, they reproduce what they learned.
On June 22, 2026, UN Women published a formal warning about the consequences of this pattern — timed specifically to precede two major UN-linked AI governance events in Geneva in early July: the UN Global Dialogue on Artificial Intelligence Governance and the AI for Good Global Summit. The warning is backed by data. And the data is more specific than most public conversations about AI and gender tend to be.
A study of 133 AI systems found that 44 per cent demonstrated gender bias. More than a quarter — 26 per cent — showed both gender and racial bias. Only 51 per cent of marketers currently use human oversight to test AI-generated creative content before it is released. And large language models have been found to consistently associate women with the words "home," "family," and "children," while associating men with "business," "executive," "salary," and "career." When tasked with completing sentences beginning with a person's gender, approximately 20 per cent of LLM responses exhibited sexist and misogynistic attitudes — including portrayals of women as sex objects and as property.
These are not edge cases or extreme examples. They are the measured outputs of the AI systems that are already shaping how billions of people see the world.
What the Bias Actually Looks Like — Across Sectors
The gender bias that UN Women is documenting is not confined to a single category of AI application. It shows up across the specific domains where AI is being deployed most aggressively and with the most consequential outcomes.
In hiring, AI screening systems trained on historical employment data reproduce historical hiring patterns. Historical hiring patterns favoured men for professional and leadership roles and women for administrative and support roles. An AI that learns from that history will recommend candidates in patterns that reflect it, regardless of the actual qualifications of the individuals being assessed. The bias is not written into any rule. It is embedded in the training data and propagated through the model's learned associations.
In healthcare, the consequences of AI gender bias are potentially life-threatening. UN Women's documentation of this issue, developed in partnership with responsible AI researchers, identifies a specific and serious gap: AI diagnostic and medical decision support tools are frequently trained on datasets that overrepresent male patients or male clinical presentations of disease. Women experience many conditions differently from men — in their symptoms, their timelines, their physiological responses to treatment. An AI system that does not know this, because its training data did not adequately represent women's health experiences, produces outputs that are less accurate for women than for men. In critical areas like healthcare, AI may focus more on male symptoms, leading to misdiagnoses or inadequate treatment for women.
In advertising and media, where AI is now used by 88 per cent of agencies in the United Kingdom alone, the stereotypes reproduced by AI-generated content are shaping how billions of people encounter gender representations in everyday life. An AI that associates women with domestic settings and men with professional ones does not need to be explicitly instructed to produce stereotypical advertising. It produces it automatically, because that is what the training data contained.
In online platforms, the intersection of AI bias and algorithmic amplification creates conditions for targeted harm. Women human rights defenders, activists, and journalists are reporting experiences of manipulated images and deepfake content at increasing rates. AI-generated non-consensual intimate imagery — deepfakes — is one of the fastest-growing categories of online violence against women. The AI infrastructure that makes it possible to create convincing fake images of any face is also the infrastructure that is being weaponised specifically against women whose public visibility makes them targets.
Why This Is Not a Technical Problem
The critical framing in UN Women's June 2026 warning is that AI gender bias is not a glitch. It is a pattern documented across systems at scale, and the pattern reflects causes that go well beyond any individual system's design choices.
The gender digital divide is the upstream problem. In low-income countries, only 20 per cent of women are connected to the internet, compared to significantly higher rates for men. The data that AI systems train on is generated by internet users. Fewer women online means less data reflecting women's experiences, women's language, women's perspectives, and women's self-representation. The training data gap directly produces the output bias.
The workforce gap compounds this. Women represent approximately 30 per cent of the global AI workforce. The decisions about what data to use, how to design models, what outputs to test for and prioritise, and what harms to guard against are predominantly being made by men. The people building the systems are not the people most affected by the systems' biases. That disconnect is structural and it is consequential.
The governance gap makes both of these harder to address. AI development is moving faster than the regulatory frameworks designed to govern it. The EU AI Act is in implementation. The US approach to AI governance remains fragmented. International coordination through bodies like the UN is essential precisely because the most consequential AI systems are global, and biases baked into them have global reach. The Geneva AI governance events in July are explicitly intended to build the kind of international consensus that can translate into enforceable standards.
UN Women's position, stated clearly in the June 22 warning, is that AI bias is not only a system design problem but also a policy problem. Technical solutions — more diverse training data, more inclusive development teams, more rigorous bias testing before deployment — are necessary but not sufficient. They need to be mandated and enforced, not left to the voluntary commitments of organisations with commercial incentives to move fast and deploy widely.
The Business Case — Why Getting This Wrong Is Commercially Costly
UN Women's June 2026 warning makes a point that is often absent from gender equity arguments addressed to corporations: addressing this is not only a matter of rights. It is commercially rational.
Research by the Stereotype Alliance, a UN Women-convened initiative, found that advertising free from gender stereotypes delivers stronger business results than stereotyped advertising. Brands using inclusive advertising recorded higher sales growth, greater customer loyalty, and stronger pricing power than competitors who did not. As AI increasingly shapes marketing and content creation, organisations that build inclusion into their AI processes are likely to benefit commercially, while those that fail to do so face reputational and commercial risk.
The Unstereotype Alliance playbook, launched in June 2026, gives marketers a specific, practical mechanism for catching bias before AI-generated content ships — an audit-and-review framework that can be integrated into the content production workflow at the point where AI is used to generate or edit creative.
The 51 per cent statistic — the share of marketers who currently use human oversight to test AI-generated creative before release — is the most immediately actionable number in UN Women's warning. It means that 49 per cent of marketers are releasing AI-generated content without testing it for the kind of gender and racial bias that UN Women's research has documented in 44 per cent of AI systems. That gap between the known bias rate and the actual oversight rate is where the most concentrated commercial and reputational risk lives.
What UN Women Is Calling For
The policy asks that UN Women is bringing to the Geneva AI governance events are specific and structural.
Gender equality must be embedded across AI development, deployment, and governance — not treated as an afterthought or a compliance checklist item but built into the design process from the earliest stages of system development. This means diverse and representative training data. It means development teams that include women in technical and decision-making roles in proportions that reflect women's share of the population and the affected user base. It means mandatory bias testing before deployment for systems with consequential outputs in domains like healthcare, hiring, and credit.
It also means legal and regulatory frameworks that treat AI-generated content facilitating violence against women — deepfakes, non-consensual intimate imagery, coordinated online abuse — as the serious harm it is, rather than leaving enforcement to voluntary platform policies that have been demonstrably inadequate.
When developed responsibly, artificial intelligence can be a powerful tool for gender equality — accelerating access to healthcare, education, financial services, and information for women who have historically been excluded from those resources. That potential is real. It is not, however, automatic. It requires deliberate choices at every stage of the AI development and deployment lifecycle.
The 133-system study, the 44 per cent gender bias rate, the 26 per cent combined gender and racial bias rate, the 20 per cent of LLM sentence completions exhibiting misogynistic attitudes — these are the current baseline. They are the outputs of systems built on decades of unequal representation. They are not the outputs of systems that were designed with gender equality as a core requirement.
Designing systems that way is possible. The Geneva AI governance meetings in July are one mechanism through which the international community can decide whether it is also required.
The AI shaping how billions of people see the world is getting women wrong. The data is unambiguous. The question now is whether the people who build, deploy, and govern that AI will decide that unambiguous is enough to act on.
