No labels, no known right answer
In supervised learning, every training example comes with an answer attached. You show the model thousands of emails, each already marked spam or not spam, and it learns to predict that label on new emails. The model always has something to check itself against, a ground truth it's trying to get closer to. Unsupervised learning removes that entirely. You hand the algorithm a pile of raw data, say, ten years of customer purchase histories, with no labels, no target column, no 'correct' output for it to aim at.
That absence changes the whole nature of the task. Instead of learning a mapping from input to a known output, the algorithm has to find structure that already exists in the data on its own: groups of similar points, dimensions along which the data varies most, points that don't fit the pattern. Nobody told it in advance what that structure should look like. This makes unsupervised learning more open-ended than supervised learning, but also harder to grade, since there's no answer key to compare against.