What is bias?

AI tools make choices all day: which video plays next, which email lands in spam, which photo gets tagged with a friend’s name. Most of the time they seem helpful. But sometimes an AI tool works better for some people than for others. When that happens in a way that isn’t fair, we call it bias.

Bias, in plain words

Bias (show definition) is a pattern that unfairly favors or works against some people or groups. People can be biased without meaning to be. So can the things people build.

An AI tool doesn’t have opinions. It learns by studying lots of examples, called training data (show definition). If those examples leave some people out, or carry unfair patterns from the past, the tool can learn those patterns too.

A simple example

Imagine teaching an AI tool to recognize dogs using thousands of photos. If almost every photo shows a big dog outdoors, the tool might struggle with a tiny dog on a couch. It isn’t ignoring small dogs on purpose. It just never saw enough of them.

Now swap dogs for people. If a tool that recognizes faces learns mostly from photos of some groups of people, it can make more mistakes for everyone else. Researchers have reported this kind of problem in real face recognition systems.¹

In the next step, you’ll look at where bias can sneak in.

Sources

1. Ketan Kotwal and Sébastien Marcel, “Review of Demographic Fairness in Face Recognition,” IEEE Transactions on Biometrics, Behavior, and Identity Science, 2025, arxiv.org/abs/2502.02309. See also Vítor Albiero and others, “Gendered Differences in Face Recognition Accuracy Explained by Hairstyles, Makeup, and Facial Morphology,” IEEE Transactions on Information Forensics and Security, 2021, arxiv.org/abs/2112.14656. ↩