Unsupervised learning trains a model on data with no correct answers attached at all, a genuinely different setup from supervised learning's labelled examples. Instead of learning to match a known, given answer, an unsupervised model has to find its own structure within the data itself, clustering, grouping similar data points together, and dimensionality reduction, compressing many separately measured variables down to the far fewer that actually matter, are the two most common tasks that fall under unsupervised learning.
Clustering finds its own groups without any predefined categories to match against
Clustering groups data points together by similarity with no predefined categories given to the model in advance, the model itself has to discover how many meaningful groups genuinely exist in the data and where the boundaries between them actually fall. That's exactly the right tool for a task like customer segmentation, where the genuinely correct groupings aren't known ahead of time and are precisely what the analysis is trying to discover in the first place.
Dimensionality reduction compresses many variables down to the few that actually matter
Dimensionality reduction compresses data originally described by many separate variables down into a much smaller number that still captures most of the data's meaningful variation. That's useful both for making genuinely huge, unwieldy datasets more manageable to actually work with directly, and for revealing real underlying structure that simply isn't obvious when looking at dozens of separate original variables all at once, unreduced.
What we're still unsure about
That clustering and dimensionality reduction are the two most common unsupervised learning tasks, working without labelled training data, is well established, uncontroversial machine learning. What's more genuinely a harder, still-debated problem is that because unsupervised learning has no correct answer to check its own results against, unlike supervised learning where accuracy against known labels gives a fairly objective measure, evaluating whether a specific clustering or dimensionality reduction result is actually good is a genuinely more subjective problem, different evaluation metrics can disagree with each other about which result is actually better, and there's no single, universally agreed way to settle that kind of disagreement the way labelled test data settles it for supervised learning.
This sits inside Unsupervised Learning (Clustering, Dimensionality Reduction), one of eight topics in AI/ML, one of seven domains in Computer Science, one of seventeen subjects the app can quiz you on.