Bias: when the data isn't fair

When the data isn't fair, the AI isn't fair either. This is called bias.

New word: Bias

Bias in A I means it treats some people or things unfairly, or works better for some groups than for others. It often happens because the data it learned from wasn't balanced. In 2018, the researcher Joy Buolamwini tested face analysis systems from big technology companies. She found they made far more mistakes with darker skinned women than with lighter skinned men, because they had learned from far fewer faces like hers.

Where bias can hide

In who is in the data. Some groups may be missing, or very rare. In who labelled it. People can add their own assumptions. In history. If past decisions were unfair, an A I that learns from them can copy that unfairness. And in who it was tested on. If it was only tested on some people, nobody notices when it fails others.

Making AI fairer

First, collect data from many different kinds of people. Second, test it with different groups, and compare the results. Third, build teams with people from different backgrounds. And fourth, let people question decisions made by A I, and get a human to look again. This is one reason it matters that Africans help build A I. When African languages, faces and voices are included, A I works better for everyone here.

Quick check

Quick check. What often causes A I bias? Unbalanced training data, where some groups are missing or rare. Options: Unbalanced training data; The colour of the computer; Fast internet; Too many charger cables.

Recap

Bias means unfair results for some groups. It often comes from unbalanced data. Testing with many kinds of people helps catch it. And diverse data and teams make A I fairer.

I've read this lesson