At Rise of AI 2025 in Berlin, Eva Gengler, co-founder of the feminist AI community, argues that today's AI is not feminist and that it must become so to avoid reinforcing existing power structures. She researches the intersection of AI and power, focusing on regulation and governance. She presents examples of bias, including AI-generated images of the Bavarian parliament that showed no women despite the body being 27 percent female, the Apple Card giving some women only 5 percent of their spouse's credit limit, an Amazon recruiting AI that automatically downgraded women, and her own study of text-to-image systems finding only 21 percent of generated people were women, dropping to 6-12 percent for 'powerful' prompts. She also cites AI used for oppression, including hijab-compliance surveillance in Iran and deepfake pornography targeting women over 90 percent of the time. Gengler presents a framework linking patriarchal, neocolonial and capitalist political context, organizational power imbalances, biased data, resource-constrained design and an automating purpose to biased AI outcomes, and proposes feminist, intersectional interventions at each stage, including diversity committees, better AI governance, and regulation, while stressing that developers alone lack the power to fix this and decision-makers must take responsibility.
Key takeaways
- Eva Gengler's own study found only 21 percent of AI-generated people were women overall, dropping to 6-12 percent when prompted for 'powerful' people.
- An Amazon recruiting AI, the subject of Gengler's master's thesis, automatically downgraded applications from women.
- Gengler argues feminism in AI is a critical, intersectional framework about power, not only about women.
- A large meta-analysis cited by Gengler found women use generative AI about 20 percent less than men, risking systems that fit men better over time.
- Gengler says developers alone lack power to fix biased AI; responsibility lies with organizational decision-makers who choose where and why AI is deployed.
Detailed notes
Why present-day AI is not feminist
Gengler opens with an AI-generated image of 'powerful people at a table' that shows only white men with near-identical hairstyles and suits, which she uses to frame her research on the intersection of AI and power, mainly on regulation and governance. She cites AI-generated images of the Bavarian parliament, which is 27 percent women, that depicted no women at all, and describes documented harms: the Apple Card giving some women only 5 percent of their spouse's shared credit limit, and an Amazon recruiting AI, the subject of her master's thesis, that automatically downgraded applications from women. She also notes AI underperforming at detecting skin cancer on darker skin and producing biased outputs in recommendation letters and image generation.
Measuring bias in image generation
Gengler describes her own study of text-to-image systems across three prompt categories: power, success and beauty. Overall only 21 percent of generated people were women, falling to 12 percent for 'powerful' and 20 percent for 'successful' prompts but rising to almost 60 percent for 'beautiful' prompts; DALL-E generated only 6 percent women for power prompts, while Midjourney generated 92 percent women for beauty prompts. She also found people of color were consistently underrepresented across systems and prompts, alongside a class bias toward thin, conventionally attractive, visibly privileged-looking people. She separately notes AI used for active oppression, including surveillance for hijab-compliance in Iran and deepfake pornography, over 90 percent of victims women, used for revenge or blackmail, plus AI-driven misinformation with political effects.
A framework linking power, context and AI outcomes
Gengler presents a framework showing how patriarchal, neocolonial and capitalist power structures shape the political and organizational context in which AI is built, for example ChatGPT's toxic-data cleaning being outsourced to low-paid Kenyan workers, and profit-first incentives within companies. This context shapes who develops and decides on AI (with little diversity among leaders and investors), what data is used (gender gaps, missing representation, stereotypes), how design trade-offs are made (developers reporting they knew data was imperfect but were denied time to fix it), and ultimately AI's purpose, which she says is currently mostly to automate existing, inequitable processes rather than change them. She adds that women both build AI less (citing a gender gap especially in Germany) and use generative AI 20 percent less than men, according to a large meta-analysis, which she says risks systems fitting men better over time.
What a feminist approach changes
Gengler defines feminism here as a critical, intersectional framework, not only about women, that examines power and the intersections of discrimination such as racism and sexism, and is activist and scientific as well as technical. Since power itself cannot be changed quickly, she argues the context of AI development must be changed instead: at the political level through equity-focused AI strategy, more funded research and stronger regulation building on the EU AI Act; at the organizational level through feminist leadership, clear AI values and governance beyond bare compliance; and at the process level through diversity committees including outside stakeholders, deliberate data correction, and governance that makes fixing biased data non-optional rather than a casualty of time pressure.
Q&A highlights
Asked why companies are not changing more, Gengler says they often simply do not have to, which is why regulation like the EU AI Act matters; she cites Gemini's 2024 controversy, where efforts to diversify generated images backfired when applied to historical prompts such as German World War Two soldiers, as evidence that some companies are trying but clumsily. Asked what individual developers should do, she says developers alone lack real power over AI systems and should educate themselves on bias, but that responsibility ultimately lies with decision-makers who often do not realize how their non-technical choices, such as where and why AI is deployed, shape outcomes.
Quotes
The future of AI must, should and will be feminist.
Eva Gengler
Developers aren't the people that really have the power in the system.
Eva Gengler
Questions answered in this talk
What does it mean for AI to be 'feminist' according to Eva Gengler?
According to Eva Gengler at Rise of AI 2025, feminist AI means applying a critical, intersectional framework, not only about women, that examines patriarchal, racist and other power structures shaping AI's political and organizational context, its data, design and purpose.
What evidence of gender bias in AI did Eva Gengler present?
According to Eva Gengler at Rise of AI 2025, examples include an Amazon recruiting AI that downgraded women applicants, the Apple Card giving some women only 5 percent of a shared spouse's credit limit, and her own study finding only 21 percent of AI-generated people were women, dropping to 6-12 percent for 'powerful' prompts.
Why does Eva Gengler say regulation is necessary rather than relying on companies to self-correct AI bias?
According to Eva Gengler at Rise of AI 2025, companies often address bias only when required to, which is why she welcomes the EU AI Act, while noting attempts at self-correction, such as Gemini's 2024 diversity changes, have sometimes backfired.