Inducing bias is (even) simpler than you think

As machine learning systems become increasingly pervasive in various aspects of our lives, concerns about the emergence of biases that disproportionately impact vulnerable populations have grown. This issue is of paramount importance, leading governments to take regulatory steps, such as the recent approval of the Artificial Intelligence Act. However, the enactment of such regulations faces significant challenges due to the complex nature of bias generation throughout the machine-learning pipeline. This complexity makes it difficult to pinpoint the root causes of bias. In this presentation, I will use tools from statistical physics to isolate and investigate one specific aspect of the problem: bias evolution across training. This aspect has been often oversimplified in the literature that aims at proposing mitigation strategies for data imbalance and fairness. I will show in a simple model that bias evolution actually presents a rich phenomenology that, depending on the statistical properties of the dataset, can lead to a catastrophic amplification of bias in naive algorithms that rely on these simplifications. This result calls for careful reconsideration in the assumptions behind heuristic methods proposed in the literature.