What is AI Bias?
AI bias occurs when machine learning systems produce systematically unfair outcomes for certain groups. Biases can arise from training data, model design, or deployment context.
Sources of Bias
Where bias enters AI systems.
Training Data
Historical biases in the data are learned by the model.
Label Bias
Human annotators introduce their own biases.
Selection Bias
Training data doesn't represent the deployment population.
Measurement Bias
Proxies used for measurement encode bias.
Types of Bias
Common categories of bias in AI systems.
Stereotyping
Reinforcing harmful stereotypes about groups.
Erasure
Underrepresenting or ignoring certain groups.
Disparate Impact
Different outcomes for different groups.
Mitigation Strategies
Approaches to reduce bias.
Diverse Data
Ensure training data represents all relevant groups.
Bias Auditing
Systematically test for bias across demographics.
Fairness Constraints
Incorporate fairness metrics into training.
Group error-rate explorer
Inspect a constructed dataset and the consequences of a shared threshold.
Compare error rates, not a “fairness score”
Constructed dataset: two groups, six cases each. Scores and reference outcomes are visible. Move the shared decision threshold and compare errors. The labels are assumed correct here; real label quality and the choice of groups require separate scrutiny.
Group A
| Score | Reference positive | Selected |
|---|---|---|
| 90 | Yes | Yes |
| 70 | Yes | Yes |
| 40 | Yes | No |
| 60 | No | Yes |
| 30 | No | No |
| 10 | No | No |
- True-positive rate
- 2/3 = 67%
- False-positive rate
- 1/3 = 33%
- Selection rate
- 3/6 = 50%
Group B
| Score | Reference positive | Selected |
|---|---|---|
| 80 | Yes | Yes |
| 50 | Yes | Yes |
| 20 | Yes | No |
| 70 | No | Yes |
| 40 | No | No |
| 10 | No | No |
- True-positive rate
- 2/3 = 67%
- False-positive rate
- 1/3 = 33%
- Selection rate
- 3/6 = 50%
TPR = selected positives / all reference positives. FPR = selected negatives / all reference negatives. Selection rate = selected / all cases. No single rate determines whether a decision system is fair; examine purpose, harms, uncertainty and data collection.
Key Takeaways
- 1Bias is often inherited from training data
- 2Different fairness metrics can conflict—choose carefully
- 3Regular auditing is essential for deployed systems
- 4Bias mitigation is an ongoing process, not a one-time fix