Ground truth contamination occurs when the labels used for training are themselves the product of a biased process that the AI system is intended to replicate or improve upon. This is distinct from annotation bias; it concerns the systemic nature of the outcomes the labels represent.

In a criminal justice context, the label “re-offended” may reflect differential policing rather than differential behaviour: communities that are policed more heavily generate more arrests, which inflates the apparent re-offence rate for those communities. In a healthcare context, labels derived from historical treatment decisions may reflect access disparities rather than clinical need. In a credit scoring context, historical default data reflects the outcomes of previous credit policies, which may themselves have been discriminatory.

The assessment examines the process that generated the labels. Was the process subject to human discretion that could introduce bias? Were there structural factors (differential enforcement, access disparities, historical discrimination) that could contaminate the labels? The Technical SME documents the label generation process, the known or suspected sources of contamination, and their potential impact on the model.

Where ground truth contamination is identified, the AISDP documents the compensating controls. These may include using proxy labels that are less susceptible to human bias (though proxy labels carry their own risks), applying bias-aware label smoothing, training the model with fairness constraints that explicitly counteract the known bias direction, or excluding the most contaminated data segments and compensating through synthetic augmentation. Where contamination cannot be adequately mitigated, this is recorded as a residual risk and communicated to the AI Governance Lead for acceptance.

Key outputs

• Ground truth contamination assessment
• Label generation process documentation
• Compensating controls for identified contamination