Bias in AI Systems

Core Contribution

Examination of how bias manifests in AI systems, from training data through deployment. Provides taxonomy of bias types and mitigation strategies.

Bias Types

Historical Bias: Training data reflects past inequities Representation Bias: Underrepresentation of groups in data Measurement Bias: Features proxy for protected attributes Aggregation Bias: Single model for heterogeneous populations Evaluation Bias: Benchmark datasets not representative

Mitigation Approaches

• Data augmentation and resampling
• Fairness constraints in training
• Post-hoc calibration
• Diverse evaluation sets

Ongoing Challenge

Bias mitigation requires tradeoffs. Optimizing for one fairness metric may worsen another. Context-dependent decisions required.

Related: 05-atom—evaluation-metric-limitations, 04-molecule—data-cascades-concept