Virginia Eubanks’ book Automating Inequality (2018) is one of the most important works of applied feminist analysis of AI published in recent years. Eubanks, a political scientist and activist, spent years embedded in communities in Indiana, Los Angeles, and Pennsylvania, documenting the impact of automated decision-making systems on poor families, particularly single mothers and their children.

Her central argument is that digital welfare systems function as what she calls a “digital poorhouse”, automating the surveillance, punishment, and exclusion of the poor in ways that reproduce and extend the functions of the physical poorhouse of the nineteenth century.

In Indiana, an automated system introduced to manage Medicaid eligibility denied benefits to hundreds of thousands of people through algorithmic errors, many of them single mothers whose children subsequently went without healthcare. The system made decisions at speed and scale that human caseworkers could not have made, and the errors it produced were proportionally greater and harder to contest. The automation did not reduce the cruelty of welfare policy. It accelerated and depersonalised it.

In Los Angeles, an algorithm used to rank homeless families for housing assistance was built on assumptions about what constituted vulnerability, assumptions that systematically disadvantaged families whose patterns of homelessness did not fit the model. Families with children, disproportionately women, scored lower than the algorithm’s designers intended and waited longer for housing as a result.

In Allegheny County, Pennsylvania, a child welfare algorithm, the Allegheny Family Screening Tool, scores families for risk of child abuse and neglect. Eubanks documents how the tool draws on data from public benefits systems, Medicaid usage, food stamp receipt, childcare subsidies, that are more complete for poor families than for wealthy ones. The result is that poor families are more legible to the algorithm and therefore score higher risk, independent of actual child welfare outcomes. The surveillance of poverty is encoded as evidence of danger.

Eubanks’ feminist analysis draws attention to the gendered dimension of these systems. The people most affected by automated welfare systems are overwhelmingly women, primarily single mothers. The systems are designed, managed, and regulated predominantly by men. The values embedded in the systems, efficiency, fraud detection, cost minimisation, reflect priorities set without meaningful input from the families they govern.

Her conclusion is not that technology is inherently oppressive. It is that technology deployed in contexts of existing inequality tends to amplify that inequality, and that the people bearing the cost of algorithmic error are rarely the people in a position to contest it.

Reflection question: Who has the power to contest algorithmic decisions that affect them in your context? What would meaningful contestability look like for people with the least power?