- Amazon’s recruiting engine, scrapped in 2018, systematically downgraded résumés containing the word “women’s”, trained on a decade of male-skewed hiring data that encoded historical exclusion into automated shortlisting.
- Sensity AI has reported that women are targeted in the vast majority of deepfake non-consensual imagery cases, with harassment and blackmail driving many to withdraw from public platforms entirely.
- A 2019 UNESCO report found that AI voice assistants including Siri and Alexa were programmed to respond deferentially to abusive language, linking this design pattern to broader normalisation of female subservience in public-facing technology.
Amazon’s hiring algorithm didn’t set out to discriminate against women, it learned to. Trained on a decade of male-dominated hiring outcomes, the tool penalised résumés that included the word “women’s” before the company scrapped it in 2018. That case has since become a reference point for a wider pattern: AI systems trained on historically skewed data don’t just reflect existing disparities, they scale them.
Hiring Algorithms and the Bias Loop
Amazon’s recruiting engine is the most documented example, but the conditions that produced it remain common. AI hiring tools are typically trained on historical hiring and performance data accumulated over decades, data that reflects who got hired, not necessarily who was best qualified. When men have historically dominated successful hires in a given role, the algorithm learns male-associated patterns as markers of a strong candidate. The result is a bias loop: past exclusion becomes present exclusion, at scale and with a veneer of objectivity.
The problem runs deeper than obvious gender markers. Algorithms can identify proxy variables that correlate with gender, certain degree types, career gap patterns, specific extracurricular language, and use those as filters without any explicit instruction to do so. Qualified women get screened out not because of anything they’ve done wrong, but because their profiles don’t match the historically male-skewed template the system was rewarded for selecting. As data integration problems continue to trip up enterprise AI deployments, the sourcing and auditing of training data for hiring tools remains an afterthought for many organisations.
Deepfakes as a Harassment Weapon
Deepfake technology has added a category of harm that didn’t exist a decade ago. Sensity AI has reported that women are targeted in the vast majority of non-consensual deepfake imagery cases online. The attacks don’t require any original intimate material, perpetrators can generate explicit content from public photographs, which broadens the pool of potential victims to essentially any woman with a public online presence.
The motivations are typically coercive: harassment, blackmail, reputational destruction. Many women withdraw from public digital platforms entirely rather than continue absorbing the exposure. That withdrawal is itself a harm, a contraction of public participation driven by a technology that platforms and legislators have struggled to keep pace with. Legal frameworks around non-consensual intimate imagery exist in some jurisdictions but remain inconsistent, and enforcement is patchy even where laws are on the books.
Voice Assistants and the Subservience Default
A 2019 UNESCO report examined the design choices behind AI voice assistants including Apple’s Siri, Amazon’s Alexa and Google Assistant, finding that all three defaulted to female voices and were programmed to respond deferentially, including to abusive language. Early versions would apologise or deflect rather than push back when confronted with insults. UNESCO argued this design pattern normalises a version of female behaviour defined by accommodation and availability, reinforcing the archetype of women as support figures rather than decision-makers. The companies involved have since updated some of these defaults, though the degree of change varies by product and market.
Data Gaps in Healthcare and Finance
In healthcare, AI diagnostic tools are only as representative as the data they were trained on, and that data has historically underrepresented women. Clinical trials have long enrolled fewer female participants, and medical records often lack sex-disaggregated data. AI systems trained on this incomplete base can misread how diseases present in women, how treatments affect female physiology, or which symptoms signal serious risk. The downstream effect is diagnostic tools that work less reliably for women precisely where accurate diagnosis matters most.
Finance has a comparable problem. Credit scoring and loan eligibility models trained on historical data absorb the gender pay gap and the traditionally male-dominated profile of prior borrowers. Women can end up penalised not because of their individual financial behaviour, but because the patterns the model learned to reward were drawn from a past that systematically excluded them. Neither sector has a clean fix, the data gaps are structural, and patching a model’s outputs without addressing its training data tends to move the problem rather than solve it.
Where Content Moderation Falls Short
Automated content moderation struggles most with the forms of abuse that are most common against women online: coded language, sustained low-level hostility, sarcasm, and context-dependent harassment that stops short of explicit threats. AI moderation tools are reasonably effective at flagging slurs and direct incitement. They are much weaker at detecting the ambient harassment that makes platforms hostile for women without crossing any single bright-line rule.
The enforcement gap cuts in both directions. Under-moderation leaves persistent harassment unaddressed, giving perpetrators effective impunity. Over-aggressive moderation, also a real failure mode, can silence marginalised voices through misclassification, a problem that falls disproportionately on the same communities the tools are meant to protect. The attack surface for AI-mediated systems is broader than most platform operators currently acknowledge, and gendered harassment sits in the gap between what automated systems can reliably detect and what actually drives women off platforms.


