How AI Learns From Data

AI is only as good as what it learns from — so data is everything. To build an AI that spots spam email, you show it thousands of emails already labelled "spam" or "not spam"; it finds patterns (odd links, "you won a prize!") and can then judge new emails. This is training. Two big ideas follow. First, more and better examples usually make a better AI. Second — and this matters — AI learns whatever is in the data, including mistakes and unfairness. If the training examples are biased or wrong, the AI will be too. That's why people who build AI have to think hard about the data they use. "Garbage in, garbage out."