Labels sound like a technical detail. They're not. They're the thing that determines which kind of learning is even possible.
A label is just the answer attached to a training example. This email is spam. This image contains a tumor. This transaction is fraudulent. Supervised learning runs on labeled data — it needs those answers to learn the relationship between inputs and outputs. The more labels you have, and the more accurate they are, the better the resulting model tends to be.
The catch is that labels are expensive. Someone has to produce them. For some problems, that's manageable: if you're building a spam filter, you have years of emails that users already marked as spam. For other problems, it's a genuine bottleneck. Labeling medical images requires trained radiologists. Labeling legal documents requires lawyers. At scale, the cost of labeling can exceed the cost of everything else in the project.
That's one reason unsupervised learning exists. When you don't have labels — or when labeling at the required scale isn't practical — you can still extract value from raw data. Unsupervised methods find structure without being told what to look for: clusters of customers who behave similarly, topics that recur across thousands of documents, anomalies that don't fit the pattern of everything else. The tradeoff is that you don't get to specify what you're looking for in advance. You find out what's there.
In practice, the two approaches are often used together. A common pattern: use unsupervised learning to find structure in a large unlabeled dataset, then apply supervised learning to the smaller labeled portion, using the structure discovered in the first step to make the second step more effective. This is part of why large language models are so capable — they're pre-trained on enormous amounts of unlabeled text before being fine-tuned on labeled examples.
The question "supervised or unsupervised?" is really the question "what data do I have, and what am I trying to find out?" The answer to that shapes everything downstream.
