The dominant narrative of AI is a narrative of machine intelligence, systems that learn, reason, and decide apparently autonomously. This narrative systematically erases the human labour that makes AI possible. And that labour is overwhelmingly performed by women, by people of colour, and by workers in the global South, the most economically vulnerable workers in the world.

Data labelling. Every supervised learning system, every AI that classifies images, transcribes speech, detects sentiment, or makes recommendations, requires labelled training data. That data is labelled by human workers, typically paid on a piece-rate basis through platforms like Amazon Mechanical Turk. A 2020 study found that the majority of data labellers globally were women, many in the Philippines, India, Venezuela, and Kenya, earning wages that frequently fell below local minimum wage standards. They work in conditions of extreme surveillance, their speed, accuracy, and output continuously monitored by the platforms that employ them. They label images of violence, pornography, and trauma that damage their mental health. They are invisible to the systems they make possible.

Content moderation. Every AI content moderation system is supplemented by human content moderators, workers who review flagged content and make decisions about what stays on and what comes down. These workers are also predominantly from the global South, employed through outsourcing arrangements that keep them off the books of the large technology companies whose platforms they protect. They are exposed daily to the worst content on the internet, child abuse material, beheadings, graphic violence, and receive inadequate mental health support. Several have sued their employers. Their labour is essential to AI-powered social media. Their existence is rarely acknowledged.

Reproductive and care labour. AI economic models that forecast the impact of automation consistently fail to account for the reproductive and care labour that sustains the economy, childcare, elder care, domestic work, emotional labour. This labour is predominantly performed by women and is significantly undervalued in GDP statistics. AI automation projections that focus on formal employment while ignoring care work systematically distort who bears the economic costs of automation. The answer, consistently, is women, especially women of colour.

Socialist feminist analysis names what is happening here: the same patterns of exploitation that characterise capitalism’s relationship with feminised labour in the physical economy are being reproduced in the digital economy. The women and people of colour who clean offices, staff call centres, and care for children are the same, or analogous, to the women and people of colour who label data, moderate content, and absorb the psychic costs of AI infrastructure. The invisibility is not accidental. It is structural.

Reflection question: The next time you interact with an AI system, consider what human labour made it possible. Whose work is invisible in the system’s performance? What would change if it were visible?