Learning from answers you already know
Supervised learning gets its name from a simple setup: every training example comes as a pair, an input and the correct output that goes with it. If you're building a model to predict house prices, each row of training data isn't just square footage and location, it's square footage and location paired with the actual sale price. The model never has to guess what 'correct' looks like, because you hand it thousands of correct answers up front and ask it to find the pattern connecting input to output.
This is different from just having a pile of data. A spreadsheet of emails with no labels tells a model nothing about what spam looks like. The same spreadsheet with a spam or not-spam column attached to every row gives the model something to aim at. The 'supervision' is that label, a human (or some reliable process) already did the work of saying what the right answer is, and the model's whole job during training is to get closer and closer to reproducing those answers on its own.