More Data Can Lead Us Astray: Active Data Acquisition in the Presence of Label Bias

More data isn't always better: The paper demonstrates that collecting additional training data can actually worsen bias in AI systems when that data contains label bias—meaning the labels themselves are systematically incorrect or unfair for certain groups. This challenges the common assumption that larger datasets automatically improve model fairness.

Label bias is often overlooked: Most bias mitigation strategies focus on either constraining the algorithm or collecting more data, but they frequently ignore bias in the labels themselves. The research emphasizes that this type of bias is distinct and requires explicit attention, not just algorithmic fixes.

Fairness constraints have limits: Even when fairness constraints are applied during data collection, they may fail to address underlying problems if they rely on biased labels. The paper shows that these constraints cannot solve bias problems they weren't designed to detect.

Different types of bias need different solutions: The research highlights the importance of distinguishing between different sources of bias (algorithmic bias, sampling bias, and label bias). Deploying a model that addresses only one type while ignoring others can have unintended negative consequences for fairness.

Comprehensive bias audits are essential: Before collecting more data or applying fairness constraints, organizations should explicitly examine whether their labels themselves are biased, as neglecting this step during data collection can undermine entire fairness initiatives.