How are AI systems trained to avoid biased outputs?

AI systems are taught to be fair by learning from lots of examples and being corrected when they make mistakes.

Imagine you're helping your friend learn to sort toys into two boxes, one for cars and one for dolls. At first, your friend might always put the same kind of toy in a box, like only putting red cars in the car box. But if you point out that blue cars also go there, your friend starts to notice patterns and learns better.

That’s similar to how AI systems are trained. They start by looking at many examples, like people who got jobs or were accepted into schools, and try to find patterns. If they only see one kind of person being chosen most of the time, they might think that's the only way someone can be successful. But if we correct them with more diverse examples, maybe showing them stories about different kinds of people getting chosen, they learn to make fairer decisions.

Sometimes, we also give AI systems a "nudge" by telling them what went wrong in their choices. This is like when you tell your friend, “Hey, that blue car should go in the car box too!” Over time, they become better at sorting, just like AI becomes better at making fair choices.