What is Bias Detection

Definition

Bias detection is the process of identifying unfair or systematic disparities in AI model behaviour across groups, using fairness metrics and testing to reveal where a model treats certain populations inequitably so the bias can be understood and addressed.

« Back to Glossary Index

Reveals where a model treats groups unfairly or inequitably
Quantifies disparities using objective fairness metrics
Enables mitigation before biased models cause harm
Supports compliance with anti-discrimination requirements

Real World Example

A team runs bias detection on its lending model and finds it approves one demographic at a materially lower rate for equal qualifications, prompting investigation and rebalancing before the model is deployed.

FAQs

How is bias detected in models?

By measuring outcomes across demographic groups with fairness metrics and testing for systematic disparities.

What causes model bias?

Unrepresentative or historically biased training data, flawed objectives, or proxy features that correlate with protected attributes.

What happens after bias is detected?

Teams investigate the cause and apply mitigations such as rebalancing data, adjusting the model, or adding fairness constraints.