Learning from examples
A conventional fraud-checking program might flag every payment over a set amount. A machine-learning model instead learns from past transactions marked as fraudulent or legitimate. It uses patterns in those examples to estimate whether a new payment may be fraudulent.
Other uses include recommendations, forecasts, identifying objects in photos, and tools that generate text.
Where it can go wrong
A model can make mistakes when new situations differ from its training examples. It may work better for some groups than others, even when its overall accuracy looks good. Results can also worsen as the world changes. People need to check how it performs and plan what happens when it gets something wrong.