Bias in AI

What Bias in AI is

Bias enters through unrepresentative training data, historical inequities encoded in labels, proxy variables that stand in for protected attributes, and evaluation that only reports aggregate accuracy.

How it works

Mitigation means measuring performance by subgroup, auditing data representation, testing with counterfactual inputs, applying fairness constraints, and documenting known limitations. There is no single fairness definition that satisfies all criteria simultaneously.

Why it matters

Biased systems in hiring, lending, policing and healthcare cause concrete harm and increasingly attract legal liability.

Common uses

• →Fairness audits
• →Subgroup performance reporting
• →Model documentation