Learning from data with no answers provided. The system has to find structure by itself. A retailer might feed in customer purchase histories and let the system group shoppers into natural clusters, without anyone deciding in advance what the groups should be. It is useful when you have plenty of data but no labels, which is most data in the world.
Technical view
Common approaches include clustering, which groups similar items, and dimensionality reduction, which compresses data into fewer numbers while keeping the important variation. A closely related idea, self-supervised learning, is what actually powers modern language models: the labels are generated automatically from the data itself. Hide a word in a sentence and the answer is simply the word you hid. This gives you effectively unlimited labelled training data at zero labelling cost, which is the single biggest reason language models could be scaled so far.
Related
Part of AI Terminology Index, section: The Foundations.