AI systems do not merely reflect gender inequality in their performance. They actively reproduce and reinforce it, at scale, at speed, and with the authority of apparent objectivity.

Consider voice assistants, Siri, Alexa, Cortana, Google Assistant. All were initially deployed with female default voices. All were designed to be helpful, responsive, deferential. All were designed to accept commands without pushback. When users spoke to them dismissively or sexually, they were designed to respond neutrally or deflect gently.

The UNESCO report “I’d Blush If I Could” (2019) documented the design of voice assistants as an exercise in the encoding of feminine subservience. The assistants are named with traditionally female names. They speak in traditionally female registers, warm, helpful, accommodating. They are designed to serve without complaint. They model a particular vision of femininity, the ideal secretary, the perfect assistant, and normalise it through billions of daily interactions.

This matters not because voice assistants are causing harm in the way that biased hiring algorithms cause harm. It matters because every technology is also a cultural artefact, it teaches us something about how the world works and how it should work. Voice assistants that model feminine subservience normalise that subservience. They train users, and especially children, to expect it from women.

In hiring, AI has been shown to reproduce glass ceilings algorithmically. Amazon’s abandoned hiring algorithm is one example. Research on AI CV screening systems has consistently found that they downgrade characteristics associated with women’s career paths, career breaks for childcare, employment in female-dominated industries, leadership in volunteer or community organisations, while rewarding characteristics more common in male career paths.

In credit and finance, AI systems reproduce wealth gaps produced by gendered economic history. Women, who have historically had less access to credit, smaller savings, and shorter formal employment histories because of discriminatory policies and unpaid care work, present financial profiles that AI systems systematically rate less favourably, without any explicit consideration of gender.

In healthcare, AI diagnostic systems trained predominantly on male patient data have consistently performed less accurately on female patients. Heart disease, historically researched and modelled on male patients, is one of the most documented examples. Women’s symptoms differ from men’s, but AI diagnostic systems trained on the male-norm data do not know that.

Reflection question: Can you identify an assumption about gender embedded in an AI system or product you know? What normalises it? What would it take to question it?