Unfair Examples, Unfair AI
Imagine you are learning what a "good drawing" looks like. But every drawing your teacher shows you as "good" has one thing in common: they were all made with blue crayons. Red crayon drawings, green crayon drawings, purple crayon drawings — your teacher always says "not quite right" about those. After a while, you might start to think: good drawings must be blue. But that is not true at all! The teacher's examples were just missing something important. AI can fall into exactly this trap. When the examples it learns from are missing people, leaving out groups, or showing the world in an unfair way — the AI learns those unfair ideas as if they were facts.
What Is Bias?
There is a word for when someone — or something — has an unfair lean toward or against a group of people. That word is bias. When AI has bias, it does not treat everyone fairly. It might work great for some groups of people and work badly for others. Here is an important thing to know: AI does not choose to be biased. It is not mean. It just learned from examples that were biased — and it copied those patterns without knowing they were wrong. That is actually good news, in a way. It means bias can be fixed by fixing the examples.
Bias means an unfair lean toward or against a group of people. When AI learns from biased examples, it can become biased too — not on purpose, but because it copies the patterns it sees.
Here is a story about how bias gets into AI. A company built an AI to help doctors figure out which patients might need extra care. They trained it on records from the hospital over the last 30 years. The problem? Thirty years ago, many people in certain neighborhoods did not go to that hospital very often because they could not afford it. So those patients were barely in the records. The AI learned: patients from those neighborhoods need less care. That was wrong! Those patients just had fewer records, not fewer health needs. The AI was biased against people from those neighborhoods — and it happened because the examples left them out. A group of researchers noticed the problem, went back and fixed the examples, and retrained the AI. The new version was much more fair.
Match each word to its meaning.
Terms
Definitions
Drag terms onto their definitions, or click a term then click a definition to match.
How can you spot bias in AI? Here are some clues to watch for. Clue 1: The AI makes more mistakes for some groups of people than others. Clue 2: The AI seems to assume things about people based on where they are from, what they look like, or who they are. Clue 3: Some kinds of people or situations are almost never mentioned in the AI's answers. When you notice any of these clues, it is worth asking: were the examples fair? Did the examples include everyone?
AI does not choose to be unfair. It just learned from the examples it was given. But even though it is not the AI's fault, bias still hurts real people. That is why it is so important to fix it.
What is bias in AI?
The hospital AI was biased against patients from certain neighborhoods. What caused the bias?
Spot the Missing Examples
- Imagine you are building an AI to help pick library books that kids will enjoy.
- You have a pile of 100 book reviews to train your AI. But when you look carefully, you notice something: 90 of the reviews were written by kids aged 10 and 11. There are almost no reviews from kids aged 5, 6, or 7.
- Think about these questions and write or draw your answers:
- What might the AI learn about what makes a book enjoyable?
- Would the AI be good at picking books for younger kids? Why or why not?
- What would you do to fix the training examples?
- Who might be left out of the AI's recommendations if you do not fix it?
- Share your answers with someone. Talk about why including all kinds of examples matters.