Data generation that encodes real-world inequalities into the model can lead to which bias?

Data generation that encodes real-world inequalities into the model can lead to which bias?

• A. Proxy label bias
• B. Historical bias
• C. Sampling bias
• D. Representation bias

Answer: (B)

Data that embeds real-world inequalities into the training signals causes models to inherit those disparities. This is historical bias: the model learns from data shaped by past decisions and conditions, so it reflects and can amplify those same inequities in its predictions. For example, if historical hiring data reflect discrimination, the model will reproduce biased outcomes rather than truly objective ones. Other biases involve different mechanisms: proxy label bias arises when the target label unintentionally stands in for a sensitive attribute; sampling bias comes from non-representative data collection; and representation bias concerns how the data are encoded or which features are used. The described scenario directly ties to past conditions embedded in the data, which is why historical bias is the best fit.

Data that embeds real-world inequalities into the training signals causes models to inherit those disparities. This is historical bias: the model learns from data shaped by past decisions and conditions, so it reflects and can amplify those same inequities in its predictions. For example, if historical hiring data reflect discrimination, the model will reproduce biased outcomes rather than truly objective ones.

Other biases involve different mechanisms: proxy label bias arises when the target label unintentionally stands in for a sensitive attribute; sampling bias comes from non-representative data collection; and representation bias concerns how the data are encoded or which features are used. The described scenario directly ties to past conditions embedded in the data, which is why historical bias is the best fit.