What does it mean to live a good woman's life in Qazaqstan?

Every woman grows up inside a set of expectations about what her life should be: when to marry, how many children to have, whether and where to work, what she owes others, and what happens if she disagrees with these expectations. We want to understand how those expectations shape a woman's life across the country, which actors have a say in these decisions, and how much.

Much has been written about what is expected of women but far less about how women live through those expectations from school to marriage to children, what they give up, what they get in return, whether they end up happy with their choices, and how that differs by region and by generation.

Imagine a hundred and twenty women, each telling us her life for an hour and a half. That is about a hundred and eighty hours of talk that we will turn into larger stories of what it's like to be a woman in different corners of our country.

The first step is transcribing and de-identifying the data, with every name and place removed, and writing each life up as a short case summary. Then we build a matrix of a hundred and twenty lives to find the patterns single interviews cannot show.

That matrix becomes the paper where we ask: at each turning point in a woman's life in Qazaqstan, how are these decisions made, which actors are present, what does it cost to disagree with them, and what are the outcomes of these choices? There also will be the book. The book matters to us as much as the papers: it is where the lives are told whole and not fitted into columns. It will be narrative nonfiction: a drive across the country, with the road, the participants, and the four of us as the thread.

Two other papers are about the method rather than the findings: one on how we worked with AI from the first draft of the design to the final book, and one on a way of reading the transcripts with mathematics rather than eyes.

This study was co-designed with Anthropic's AI model, Claude, from the first day. Claude has been a fifth member of the team at every stage: research co-design and adversarial testing, the interview guide draft, the sampling and recruitment strategy, the analysis plan, and, yes, even this page.

Following our ethics protocol, approved by Nazarbayev University's research ethics committee, we are committed to explaining to every participant how their data will be used. The audio recordings stay only on our own computers, where we transcribe them. Then every name, place and detail that could point to participants is removed. Only that de-identified transcript is shared with Claude, and every participant is aware of this from the consent form.

Why bring an AI model in at all? Because we want to see how AI can change the analysis of qualitative data. One part we are seriously curious about is text embeddings for our data. A language model turns each written passage into a long list of numbers, a position in a space with hundreds of dimensions, where passages that mean similar things sit close together. That holds across languages: one woman's answer in Qazaq and another's in Russian can land side by side if they say the same thing, with no translation in between. Once every answer to a question like “what do people here think a woman should do” is a point in that space, you can ask the space questions. Which answers cluster? How far apart do “what people here think” and “what I think myself” sit for each woman? Which women sit between the groups the human readers built? We do not give the model our categories. We ask what groups it finds on its own, then compare them with ours. Whether this works for Qazaq is itself a test: we check the model against a simple word-count baseline and a public Qazaq and Russian parallel corpus, and we check that it is not just sorting participants by language. The model runs on our own laptops.