In machine learning theory, model bias is the error that comes from a model’s simplifying assumptions, which cause it to consistently miss the true pattern in the data. A high-bias model is too simple and underfits. The term “bias” is also widely used for fairness bias: when a model produces systematically unfair outcomes for certain groups. This page explains both, because they are often confused.
Key takeaways
- Statistical bias is systematic error from overly simple models.
- It trades off against variance: reducing one often increases the other.
- Fairness bias is about unequal or discriminatory outcomes, usually rooted in data and design choices.
- A model can have low statistical bias and still be highly unfair.
Statistical bias explained in simple terms
Imagine an archer whose arrows always land in a tight group, but consistently to the left of the bullseye. Their aim is consistent but systematically off. That is high bias. In a model, high bias means predictions are consistently wrong in the same direction because the model’s assumptions do not match reality, such as fitting a straight line to a curved relationship.
The bias-variance trade-off
A model’s expected prediction error can be split into three parts: bias squared (error from wrong assumptions), variance (error from sensitivity to the specific training data) and irreducible error (noise no model can remove).
| Model state | Bias | Variance | Result |
|---|---|---|---|
| Too simple | High | Low | Underfitting |
| Balanced | Moderate | Moderate | Best generalisation |
| Too complex | Low | High | Overfitting |
As model complexity increases, bias usually falls and variance rises. The best model sits where total error is lowest.
How to reduce statistical bias
- Use a more flexible model, such as gradient boosting instead of linear regression.
- Add informative features through feature engineering.
- Reduce regularisation strength.
- Train longer, if training stopped too early.
Fairness bias: the other meaning
In everyday and policy language, “AI bias” or “model bias” usually means unfair outcomes. For example, a hiring model that ranks equally qualified candidates differently by gender, or a face recognition system with higher error rates for some groups.
Common sources of fairness bias
- Historical bias: the data reflects past discrimination, such as past hiring decisions.
- Sampling or representation bias: some groups are under-represented in the training data.
- Measurement bias: labels or features are recorded differently across groups.
- Proxy variables: features like postcode can stand in for protected traits.
- Objective choice: optimising only overall accuracy can ignore poor results for small groups.
How to reduce fairness bias
- Audit data coverage across groups before training.
- Measure error rates by group on the test set, not just overall accuracy.
- Remove or test proxy features.
- Apply fairness techniques such as reweighting, constraints or group-specific thresholds where appropriate and lawful.
- Keep human review and appeal routes for decisions about people.
- Document limits using tools like model cards.
Statistical bias vs fairness bias
Statistical bias is about accuracy: is the model systematically wrong? Fairness bias is about ethics and law: are some groups treated unfairly? They are separate. A highly accurate, low-bias model can faithfully learn and repeat unfair historical patterns. Fixing one does not automatically fix the other.
Why it matters
High statistical bias wastes data and gives poor predictions. Fairness bias can harm people, break anti-discrimination laws and damage trust. Regulations such as the EU AI Act require risk management and data governance for high-risk systems, including attention to bias.