In 2018, a survey of AI researchers at major conferences found that women made up approximately 12% of participants. In machine learning specifically, the technical field most directly responsible for contemporary AI systems, the figure was similar. Among the small number of Black and Latina women researchers, representation fell to fractions of a percent.

These are not merely diversity statistics. They are structural facts about who makes the decisions that shape AI systems. And feminist theory is clear that homogeneous groups, groups drawn from similar backgrounds, with similar experiences, holding similar assumptions, produce systematically limited outputs. Not because individuals within those groups are malicious or incompetent, but because the perspectives they bring reflect a narrow slice of human experience.

The consequences are concrete and documented.

Voice recognition systems trained predominantly on male speech patterns performed poorly on female voices when first deployed commercially. The problem was not that the engineers hated women. It was that the voices they tested the systems on were predominantly theirs, and male voices were implicitly treated as the norm.

Image recognition systems have consistently performed poorly on images of women in professional contexts, classifying female doctors as nurses, female executives as assistants. The training data reflected existing professional hierarchies rather than aspired ones.

Sentiment analysis systems have been shown to classify text associated with women’s experiences, menstruation, pregnancy, domestic violence, as negative or low-credibility, reflecting cultural stigmas encoded in the data.

Natural language processing systems have reproduced gendered stereotypes at scale. When asked to complete sentences, they have associated career terms with men and family terms with women, technical expertise with men and caring roles with women, patterns that reflect historical bias in the text corpora they were trained on.

In each case, the problem is not a single bad decision. It is a pattern of decisions made by teams that did not include, or meaningfully consult, the people whose experiences were being encoded. Feminist theory identifies this as a structural problem, not an individual one. The solution is not diversity for its own sake but the structural inclusion of diverse perspectives in design processes, with the power to shape, not merely advise.

Reflection question: Think about the last technology product you used. Whose needs were clearly anticipated in its design? Whose were not? What does that tell you about who was in the room when it was built?