AI Is Deciding Our Lives. Are Women Deciding AI?

This piece is adapted from an invited talk at the FEMINNO Armenia CSW70 parallel event, part of the United Nations Commission on the Status of Women.

Language is one of the primary ways power operates in society. As AI systems increasingly interpret human language, they shape what we see, how we are evaluated, and how information is summarized.

This raises a critical question: who is visible to these systems—and who is not?

Women are already underrepresented in the AI world. But not all women are affected equally. Women from low-resource language communities (those without a strong digital presence), and women from minority groups, face a form of double erasure ... by gender, and by language. Their ways of speaking, cultural norms, and lived realities are often absent from the data used to train AI systems.

This is not only a question of bias. It is a question of invisibility.

Many languages spoken in marginalized or indigenous communities have little to no representation in the data used to train AI systems. In some cases, they lack even basic digital resources such as corpora, benchmarks, or evaluation frameworks. When these languages are missing, the people who speak them are excluded as well.

Those absences are then carried into AI systems that shape access to critical domains such as legal services, healthcare, employment.

So we have to ask:

• Whose language counts as data?

• Whose norms are treated as standard?

There is also a deeper challenge. When communities are asked to contribute their data to AI systems, many hesitate—and with good reason. Too often, data is extracted, tools are developed elsewhere, and little value is returned. In some cases, these tools are even used in ways that harm the very communities they were meant to represent.

The answer is not simply more participation. It is ownership.

At a basic level, this means communities have control over how their language data is collected, used, and shared. In practice, this can take different forms: community-owned datasets, licensing agreements, or models trained locally so that data never leaves the community.

These approaches are not theoretical. We already see them in other domains—such as healthcare and Indigenous data governance. AI has simply not caught up.

There are also concrete technical and legal mechanisms that support this:

• Data trusts, where a community collectively owns and governs its data, deciding who can access it and under what conditions

• Federated learning, where models are trained locally and data never leaves the community

• Benefit-sharing frameworks, where value generated from data is returned to the communities that contributed it

These models shift the question from access to agency.

To see why this matters, consider a simple example. Persian has a social system called taarof, a form of politeness where what is said and what is meant are deliberately not the same. Offers are refused before being accepted; requests are often indirect. This is not ambiguity, but a structured system encoding social relationships and norms.

When we tested large language models on taarof, they struggled—not because they lacked vocabulary, but because they lacked cultural understanding.

And Persian, with over 90 million speakers, has a relatively strong digital presence.

Now consider languages with far less representation such as Kurdish, Balochi, Mazandarani, Taleshi, spoken in communities that are already marginalized, and among the least visible in AI systems.

If we only look at gender in dominant languages, we miss this layer entirely.

If we want a more equitable AI future, women cannot be only users of these systems. They must be involved in shaping the data, defining the norms, and governing how language is interpreted across cultures.