女性真正需要的 AI，並非「AI wife」

In late October, The New York Times introduced readers to Dazzle, an AI assistant built by former Yahoo CEO Marissa Mayer, with an anecdote that was meant to charm and humanize. 

When a teacher asked for three bullet points about what made her 10-year-old daughter special, Mayer turned to Dazzle. The app scanned her camera roll and reported back: empathetic, creative, an animal lover. Mayer called the answer “perfect.” 

This personal narrative was Mayer’s attempt to show what an AI tool designed by and for women looks like. Rather than require complex coding or agent setup, Dazzle will learn about you using data scraped from your digital photo album, and spark a dialogue to assist you with daily, quotidian tasks. At one point, Mayer describes Dazzle as an “AI wife” meant to offload the mental labor that disproportionately falls on women.

But those comments put Mayer at the center of a social media storm, with women and parents asking questions such as: “What kinds of domestic labor should we outsource to AI?” and, “What do women really want or need from AI?”

More from CNET

As the mother of a 9-year-old, I had a visceral negative reaction to Mayer’s example. Of all the parenting chores I’d like to outsource, a task designed to make my child feel special and unique isn’t one of them.

The teacher’s prompt was not administrative. It was an invitation to sit with a child and explore the question of who she is – an intellectual and emotional exercise that builds a 10-year-old’s confidence and self-esteem far more than any adjective the teacher receives. Mayer skipped the conversation but kept the adjectives.

Mayer has since told Fortune that she was testing the product, not outsourcing a parenting moment, and that her aim is simply to give women time. I take her at her word. But her defense concedes the critical point. A founder’s chosen example of her product, which likely went through rounds of internal PR finessing, tells you what the product is for, and this one was built to answer, on a mother’s behalf, a question only a mother or father can answer. 

The problem is not that Dazzle exists. It’s that an assistant that files school forms and remembers children’s allergies has been marketed to women as the thing we have been waiting for, when the things women have actually been waiting for are much harder to build and far less profitable to sell.

What I want from AI, and what I refuse to hand over

I ran the nonprofit AI4ALL, the entire purpose of which is to put technical skills for responsible AI in the hands of young people who have been locked out of it. I am not a skeptic. I want AI that finds errors in a contract, pulls research for a grant, analyzes reams of data and takes the drudgery out of my work to give me space for the parts that require thought and rigor.

What I won’t hand over is knowing my own child. I can think of a dozen things that make my daughter special without consulting an AI model. She has a ribald, slightly scatological sense of humor that’s disarming and hilarious. She’s deeply musical and can play most tunes by ear. She is extremely good at math, hates new math and would much rather do it the old-fashioned way. She has a strong sense of justice, is sassy, flamboyant and ridiculous. She is perfect. 

None of that is in my camera roll, and no AI model would find it there, because it lives in the hours of parental labor I have spent paying attention to her. Describing her is not a chore to be optimized. It’s the relationship itself, manifested in narrative.

I don’t want to outsource my parenting to AI, and I don’t want an AI wife. Set aside, for a moment, what it says about the tech industry that “wife” is the readiest available word for a thing that remembers the dentist and books the sleepover. Ample sociological literature examines the mental load women carry in households, and Dazzle’s premise is that AI can lift it. But the mental load isn’t a single, undifferentiated set of tasks. Some of it is logistics. Some of it is love disguised as logistics, and AI can’t separate the two. When the Times asked Mayer whether a future of AI-planned birthdays and holidays was a little depressing, she said she hadn’t really thought it through. That’s the most honest line in the splashy profile.

I want AI that sees me as a woman, not as a demographic. I want AI that is not built on centuries of gender-biased data. I want AI built on women’s knowledge of our own lives, one that doesn’t advise us to ask for less money, steer us toward lower-paying work, or render us younger and thinner than we wish to be when touching up a headshot. 

The deeper objection to Dazzle is the one Simone de Beauvoir made to a different technology of womanhood 70 years ago: Men describe the world from their own point of view and mistake it for the truth. De Beauvoir was referencing the masculine machinery of literature, science, myth, religion and law — all things that LLMs now digest and fabricate with an ease that makes many users believe the AI may be verging on consciousness. An assistant pitched at women, arriving with a built-in theory of what women need — a theory it didn’t ask us for — packaged as a product roadmap.

Mayer told the Times that women take more photos and are more visual, and called it her own bias. It is. Everything in that profile reinforces a gender stereotype rather than defying one. The pitch assumes women want AI to be a better domestic manager. What women need from AI is something else entirely.

When AI can’t ‘see’ you

In December 2023, the New York Times published a roster of the people behind modern AI. It included no women: not Fei-Fei Li, whose ImageNet made the deep-learning era possible, and not Mira Murati, then chief technology officer of OpenAI. 

The philosopher Miranda Fricker calls this epistemic injustice, the wrong done to someone in their capacity as a knower. Related is ontological erasure: It’s not enough that women’s contributions go uncredited; the category “person who builds AI” is quietly defined so that women fall outside it. 

When the Times asks whether someone like Mayer – a former Yahoo CEO, and before that a software engineer and vice president at Google – can fix AI’s gender problem, it asks a woman to solve a problem the paper’s own past framing helps sustain.

What happens when you build on biased data

The models behind products like Dazzle are trained on the internet’s record of what people have written about women, and that record isn’t kind. 

A 2024 study commissioned by UNESCO (PDF) and led by researchers at University College London tested OpenAI’s and Meta’s models and found that female names clustered with “home,” “family” and “children,” while male names clustered with “business,” “executive” and “career.” Asked to complete sentences about a woman, Meta’s Llama 2 produced sexist or misogynistic content roughly one time in five. Asked for occupations, the models gave men “doctor” and “engineer” and gave women “domestic servant,” “cook” and “prostitute.” 

Researchers at Apple reported at the 2023 ACM Conference on Fairness, Accountability and Transparency that when they challenged large language models on stereotyped occupational choices, the models invented false justifications rather than correcting them.

This is the raw material from which Mayer has built her AI assistant for women. The models that already believe women belong in the kitchen won’t need much encouragement to decide that what a busy mother wants is help planning a sleepover.

The risk compounds in three ways. First, the stereotype becomes the product: An assistant that learns from your photos what a woman like you is interested in will return you to yourself, but regress you to the mean each time. Second, it becomes the labor market. The World Economic Forum’s Global Gender Gap Report 2026 found that AI is now disproportionately automating clerical and white-collar roles that gave women economic independence, while women remain concentrated in low-paid data-annotation work within AI itself. Third, it becomes the culture. Increasingly, people are engaging with synthetic, AI-generated “women” in the form of feminized assistants, synthetic influencers and companion app “girlfriends” designed to be docile. This synthetic engagement trains people to become accustomed to a more passive engagement model, which results in greater stress when engaging with women in the real world.

None of this argues against AI. It argues that AI built on existing data, by the existing workforce, with existing capital, will deliver women a more efficient version of the world that erases them.

Here are four things that would be truly beneficial for women.

AI that isn’t biased against us. Not an assistant tuned to our stereotypes, but models audited by people outside the company for how they treat women in hiring, lending, medicine and credit. Every woman who asks an AI for a salary benchmark, a symptom check or a mortgage comparison is already using it as a tool of consequence, and the UNESCO findings mean she is doing so on a model that quietly expects less of her. 

Jobs in the industry that builds it. The WEF’s 2026 report found that women hold 19.3% of AI engineering roles and are underrepresented at every level inside AI firms, more so than in comparable non-AI companies. LinkedIn’s analysis of 15,000 companies across 27 countries, published in August, found women held 13% of AI executive positions and made up 26% of American AI hires in 2025. Mayer told the Times she wants to see more people like her at the table.

Optimism is not a plan. Hiring, promoting and retaining women across the model labs moves the needle, and so does pulling women out of the annotation farms and into the roles that decide what the annotation is for.

An end to our erasure as knowers. The pattern that lost women in the early days of computer science is still running. It runs when a newspaper’s who’s-who of AI contains no women. It runs when AI research citation networks route around the women, many of them women of color, who identified the harms first: Joy Buolamwini on facial recognition, Timnit Gebru on the eugenic undertow of the field’s ambitions, Safiya Noble on search, Ruha Benjamin on the new Jim Code.

Reversing it is not a matter of adding women to the list. It means treating the question of who gets to count as an AI builder, and whose knowledge about AI counts as expertise, as something that has been decided wrongly and must be decided again.

Capital, on our own terms. PitchBook’s 2025 All In report recorded a headline that looks like progress: a record $73.6 billion to US companies with at least one female founder, 27.7% of all venture dollars. But roughly two-thirds of that went to AI, and two companies, Scale AI and Anthropic, accounted for more than $30 billion. Strip out a handful of megadeals with a woman somewhere on the founding team and the picture for all-female teams is what it has been for a generation: low single digits. 

What would actually work

So where do we begin? Here are five immediate actions.

Ask women what they need before you build it, and let the answer change the product. Participatory design, in which the people who live with an AI model’s consequences help decide what it does, is the practice with the strongest evidence, and it’s the one most startups skip because it slows the roadmap. A women’s assistant designed by women would, I suspect, start by helping with pay transparency or medical self-advocacy rather than birthday parties.

Audit for gender before release, and publish the results. UNESCO’s 2024 study and the Recommendation on the Ethics of AI it rests on, adopted by 193 member states in 2021, call for bias audits before and after deployment and for disclosure of training-data composition. The European Union’s AI Act now requires impact assessments for high-risk systems. A company that markets to women and will not say how its model performs for women is asking for trust it has not earned.

Treat representation as an input, not a public relations output. The question is not whether a woman is CEO. It’s who was sitting in the room when the product’s consumer profiles were written, who labels the data, who decides what counts as an error. The WEF data show that the higher you go in an AI company, the fewer women there are. That is where the model’s values are actually set.

Design for autonomy, not deference. A technofeminist design principle, in the tradition of Judy Wajcman, holds that a tool should expand what its user can do and know, not manage her within an assumed role. The test is simple. Does the product make a woman more capable of acting on her own behalf, or does it make her a more efficient administrator of a life someone else designed?

Move capital, not just conversation. For years, the WEF has recommended ring-fenced funds for gender parity and financial incentives for women’s entrepreneurship. Until all-female founding teams receive more than a rounding error of venture capital, the women with the best ideas about what AI could do for women will keep losing to the men, and women, with the best access.

AI inspired by gender empowerment

I believe that more women building AI will produce more ethical, more responsible AI. But a woman at the top is not the same as a feminist at the top.

Mayer told the Times she does not see gender inequality in tech as a pressing issue, that gender “fades into the background,” and that passion is “a gender neutralizing force.” This is the same woman the article describes as being dissected by a press that couldn’t decide whether she was frivolous or too hard-driven, and then labeled her a failure.

The women who will change this industry are the ones who actually believe in gender empowerment, not the ones who have bought into the patriarchy and learned to sell it back to us in a softer voice. What women need from AI is not an assistant that knows what women are like. It’s an industry that stops assuming it already knows, employs us to build the models, credits us when we do, funds us when we start companies and audits itself when it is wrong.

Bo Young Lee

Guest columnist

Bo Young Lee is a globally recognized AI, tech and human capital ethicist, DEI and ESG executive, and public speaker. She is the former CEO and current strategic adviser at AI4All, a nonprofit helping build AI to serve humanity and society. Before that, Bo was president of research and advisory for AnitaB.org, a group advancing women and nonbinary technologists. She also was Uber's first chief diversity, equity and inclusion officer, leading a transformation of its culture, values and equity environment. Earlier, Bo was the first global diversity and inclusion officer for Marsh McLennan's risk and insurance services businesses and launched and led Aon Hewitt’s global emerging workforce solutions practice. Bo holds a 1st-level master of studies in AI ethics and society from University of Cambridge/Leverhulme Center for Future Intelligence, an MBA with distinction from NYU’s Stern School of Business, and a BBA magna cum laude from University of Michigan’s Ross School of Business. See full bio