Removing protected attributes fails to remove bias from models

Evidence

Simply removing the “protected attribute” is insufficient. As long as the model takes in features that are correlated with, say, gender or race, avoiding explicitly mentioning it will do little good. — Brian Christian, The Alignment Problem, 294

Categorization

• DEFINE: Introduces the counterintuitive core of the ‘redundant encodings’ concept.
• corroborated: 2 independent passages support this claim — Brian Christian, The Alignment Problem, 294
• XREF: connects to: Partial debiasing can make bias problem worse by hiding it — Brian Christian, The Alignment Problem, 210; Untested risk models can detain people on factors having no real predictive value — Brian Christian, The Alignment Problem, 269; Machine-learning bias often stems from unrepresentative training data

Related

• Privacy rules can prevent engineers from even measuring model bias (similarity 0.82)
• Partial debiasing can make bias problem worse by hiding it (similarity 0.82)

Broader Connections

• Machine-learning bias often stems from unrepresentative training data #expands (0.70)