What “algorithmic bias” means, and three ways to test for it

Algorithmic bias describes systematic and repeatable errors in a computer system that produce unfair outcomes, such as favouring one group of people over another. It can come from the data a system learns from, from the way a problem is defined, or from how the system is used.

Three practical checks anyone can ask about:

1. Compare outcomes across groups. Does the system's error rate, approval rate or recommendation differ by group, such as age, gender or region, in ways that cannot be explained by relevant factors?

2. Look at the data. Does the training data represent the people the system will serve, and does it encode past unfairness, for example in historical hiring or lending decisions?

3. Check the feedback loop. Does the system's output change future data? A tool that sends more police to one area will record more crime there, reinforcing the pattern.

Limits: fairness has several definitions, and they cannot all be satisfied at once, so any assessment must state which one it uses and why.

Anyone deploying such a system should publish the checks they ran, not just the claim that it is fair.