Bias and fairness
AI learns from data collected by people, so if the data is unbalanced or unfair, the AI can be unfair too. This is called bias. People must check AI and fix bias so that it treats everyone fairly.
📌 Key points
- Bias means being unfairly in favour of, or against, some people or things.
- AI learns from data. If the data is unbalanced or unfair, the AI can learn the unfairness too: unfair data → unfair results.
- Example: a face app trained mostly on photos of one group of people may work less well for other people.
- Data can be biased if some groups are missing, if there are far fewer examples of them, or if past human decisions in the data were unfair.
- AI does not choose to be unfair. It has no opinions; it copies the patterns in its data.
- To make AI fairer: use data that includes everyone, test it with many different people, and let people check and question its decisions.
- Fairness matters most when AI helps make important decisions about people.
📖 Key words
- Bias
- Being unfairly in favour of, or against, some people or things.
- Fair
- Treating everyone equally and justly.
- Unbalanced data
- Data with far more examples of some groups than others.
- Diverse data
- Data that includes many different kinds of people or things.
- Check
- When people test an AI and look at its results to find mistakes or unfairness.
✏️ Worked example
A speech app was trained on recordings of adults. Children say it often gets their words wrong. Why, and how could it be fixed?
- The training data had hardly any children's voices, so the data was unbalanced.
- The AI learned patterns from adult voices, so it works less well for children. This is bias.
- The AI did not decide to be unfair; it only learned from the data it was given.
- Fix: add many recordings of children (with permission), train it again and test it with children's voices.
- Answer: unbalanced data gave unfair results. More balanced data and careful testing make it fairer.
When you hear that an AI made a decision, ask: "Who was in the data, and who was missing?"
Test yourself8 questions from a bank of 19. A new mix every time.
Start the quiz →