Indirectly using a protected attribute due to correlation with other features is an example of which bias?

Indirectly using a protected attribute due to correlation with other features is an example of which bias?

• A. Redundant encoding
• B. Proxy label bias
• C. Sampling bias
• D. Automation bias

Answer: (A)

This question tests how a model can leak protected information through features that are correlated with that attribute. When a protected trait isn’t used directly, the model may still infer it from other inputs that carry the same signal. That’s why this is described as redundant encoding: the protected information is encoded in multiple features, so even without the direct attribute, the model ends up using it indirectly and exhibits biased behavior. Other biases don’t capture this mechanism. Sampling bias comes from how data is collected, automation bias from over-relying on automated systems, and proxy label bias relates to biases in the labels themselves rather than the input features conveying the protected information.

This question tests how a model can leak protected information through features that are correlated with that attribute. When a protected trait isn’t used directly, the model may still infer it from other inputs that carry the same signal. That’s why this is described as redundant encoding: the protected information is encoded in multiple features, so even without the direct attribute, the model ends up using it indirectly and exhibits biased behavior.

Other biases don’t capture this mechanism. Sampling bias comes from how data is collected, automation bias from over-relying on automated systems, and proxy label bias relates to biases in the labels themselves rather than the input features conveying the protected information.