Not My AI: A Feminist Framework to Challenge Public-Sector Algorithmic Decisions

Joana Varon and Paz Peña’s Not My AI is not a general guide to “bias in AI.” It begins with public services in Latin America and asks why governments are bringing algorithmic decisions into education, social benefits and child protection. The authors look beyond model accuracy: Is automation necessary? Who designed the system and whose data trained it? What public institution will act on its outputs, and can the people affected challenge them? The project develops a feminist and anti-colonial framework for questioning these systems, resisting the presentation of technological innovation as a shortcut around poverty and inequality.

The report examines Chile’s Childhood Alert System (Sistema Alerta Niñez), which uses administrative records to estimate risk among children and families and help local offices organize support. Not My AI argues that converting records from existing public services into risk rankings can expose poor families to more intensive identification and intervention; the assurance that a human makes the final decision does not prove that a score will not be treated as neutral fact. The case account also says the government commissioned an algorithmic audit and rejected disclosure of its results. The feminist question is not simply whether an algorithm replaces a social worker. It is who defines which families need intervention when welfare is organized through risk-based triage, and who can see how the categories were produced.

A second set of analysis concerns Argentina’s Social Intervention Technology Platform and a later project linked to Brazil’s social programs, Projeto Horus. The platform attempted to use family and living-condition data about girls in poor regions to predict school dropout and adolescent pregnancy. The authors and related case record question its statistical method, data bias and gender design: concentrating pregnancy prediction on girls can turn a structural issue into surveillance of young women’s bodies. When the project reached Brazil, public documentation also did not establish whether the pilot produced effective results or how a technology company would be accountable. These are critiques of historical projects and gaps in their evidence trail, not claims about every Brazilian public system today.

The anti-colonial lens is not merely a label for geography. It points to a governing relationship: people in poor communities supply highly sensitive information, while public institutions and technology vendors gain the ability to classify and predict; affected communities may have little say in defining the problem or reviewing the system. The report’s mapping was an initial, non-exhaustive exercise conducted around 2021, not a complete current inventory, and the projects differed in status and available evidence. Its questions are useful before accuracy is measured: Why automate? Could the service work without collecting more data? Who can appeal, correct a record or refuse? Read it with Feminist Data Manifest-No and Feminist Principles of the Internet — Version 2.0 to see how refusal, governance and data control reinforce one another. Sources include the FIRN research text, the Chile Childhood Alert case and the Argentina–Brazil adolescent-pregnancy prediction case.