AI bias refers to systematic and unfair discrimination in AI system outputs that disadvantage certain groups based on characteristics like race, gender, age, or socioeconomic status. Bias can enter AI systems at multiple stages: through biased training data, flawed algorithm design, biased feature selection, or prejudiced human oversight.
Types of AI bias include: Historical Bias (training data reflects past societal biases), Representation Bias (underrepresentation of certain groups in training data), Measurement Bias (proxy variables that correlate with protected characteristics), Aggregation Bias (models that don't account for subgroup differences), Evaluation Bias (testing metrics that favor certain groups), and Deployment Bias (using a model in contexts it wasn't designed for).
Real-world examples include AI hiring tools that penalized women's resumes, facial recognition systems with higher error rates for darker skin tones, healthcare algorithms that underestimated the needs of Black patients, and credit scoring models that discriminated based on zip codes.
Mitigation strategies include diverse and representative training data, bias auditing tools and fairness metrics, regular testing across demographic groups, human oversight of AI decisions, transparent documentation of model limitations, and regulatory compliance with anti-discrimination laws.
