New AI fairness paper says dropping variables stokes bias

A new arXiv paper from Ivan Luciano Danesi and colleagues tackles a core issue in AI fairness: how we choose which variables a model sees. The authors argue that traditional variable selection practices—often meant to simplify models—can silently bake in discrimination against protected subgroups. Under the recently introduced EU AI Act, which demands trustworthy and fair AI, this becomes a regulatory problem, not just an ethical one. Their approach uses a mathematical framework to evaluate fairness while keeping ethics and social context at the center, rather than treating them as afterthoughts.

The paper's central claim is counterintuitive: removing sensitive variables like race, gender, or age—a common tactic to 'blind' models to bias—can actually harm equity. When you exclude these variables, you lose the ability to detect disparate impact across subgroups. Instead, Danesi et al. suggest retaining all potentially relevant variables, which allows for more granular fairness audits and reduces implicit bias in predictions. They call for interdisciplinary teams—mathematicians, ethicists, and social scientists—working together to meet both ethical standards and the EU's compliance requirements, positioning this as a practical path toward responsible AI deployment.