What is bias in AI?
Bias in AI is a systematic error in a model’s outputs, as opposed to random noise that averages out. The word carries two meanings that often get mixed up. In statistics and machine learning, bias is the consistent gap between what a model predicts and what actually happens; a forecasting model that always underestimates demand in December is biased in this sense. In the social sense, bias is a pattern of outputs that treats groups of people unequally, such as a resume screener that ranks women lower for technical roles. The first kind is an accuracy problem. The second is a fairness problem, and it often has legal consequences.
Types of bias
| Kind | What it looks like | Where it comes from |
|---|---|---|
| Statistical bias | Predictions consistently too high or too low | A model too simple for the pattern (see bias-variance tradeoff), missing features, shifted data |
| Selection bias | Training data does not represent the population the model serves | Records only from customers who were approved, sensors only on newer machines |
| Measurement and label bias | The target variable is a poor stand-in for the real goal | Using arrests as a proxy for crime, or cost of care as a proxy for medical need |
| Social bias | Unequal outcomes across groups of people | Historical decisions encoded in data, proxies for protected traits, uneven data quality |
Why bias matters
Statistical bias silently costs money. An inventory model that underpredicts demand produces stockouts every season until someone checks its errors by month. Social bias exposes the company to discrimination claims, regulatory action, and public criticism, and it can hide behind good overall accuracy. A hospital model that predicts the need for extra care using past spending will favor patients who historically received more care, which can disadvantage groups who were undertreated. Both kinds are found the same way, by comparing predictions with outcomes, sliced by time, segment, and group.
What to watch for
- Check residuals by segment, since a model can be unbiased on average and badly biased for a region or product line.
- Question what the training labels actually measure before trusting what the model learned.
- Removing a protected attribute does not remove social bias when other features stand in for it.
- Build bias checks into model monitoring so they run after every retraining, which is where bias detection and fairness auditing fit.