AI Ethics, Bias, and Fairness
Types of Bias in ML Systems
Data Bias
| Type | Description | Example |
|---|---|---|
| Selection bias | Non-representative training data | Medical AI trained only on one demographic |
| Historical bias | Data reflects past inequities | Resume screening inheriting hiring biases |
| Measurement bias | Flawed data collection | Proxy variables encoding protected attributes |
| Label bias | Subjective or biased annotations | Annotator demographics affecting labels |
Algorithmic Bias
- Model architecture choices favoring certain patterns
- Optimization objectives not aligned with fairness
- Feedback loops amplifying biases over time
Fairness Metrics
Group Fairness
| Metric | Definition |
|---|---|
| Demographic parity | Equal positive prediction rates across groups |
| Equalized odds | Equal TPR and FPR across groups |
| Calibration | Predictions equally accurate across groups |
Individual Fairness Similar individuals should receive similar predictions.
Bias Mitigation Strategies
Pre-processing
- Data rebalancing and augmentation
- Removing or obscuring protected attributes
- Collecting more representative data
In-processing
- Adversarial debiasing during training
- Fairness constraints in objective function
- Multi-task learning with fairness objectives
Post-processing
- Threshold adjustment by group
- Calibrated predictions
- Human review for high-stakes decisions
Responsible AI Frameworks
- NIST AI Risk Management Framework
- EU AI Act requirements
- Model Cards and Datasheets
- Algorithmic Impact Assessments
Best Practices 1. Document data sources and known limitations 2. Evaluate on disaggregated metrics by protected groups 3. Include diverse perspectives in development 4. Implement ongoing monitoring for drift and bias 5. Create feedback mechanisms for affected communities
ethicsbiasfairness
Explore 500+ Semiconductor & AI Topics
From EUV lithography to CUDA optimization — search the full knowledge base or chat with our AI assistant.