Supervised learning trains a model on a labelled dataset, examples that already carry the correct answer attached, and the model's entire job during training is to adjust its own internal parameters until its predictions match those known answers closely enough on the training data. Two broad kinds of task fall under supervised learning: regression, predicting a continuous number, a house's likely price, say, and classification, predicting which discrete category something belongs to, whether an email is spam or not.
Regression predicts a number, classification predicts a category, both need labelled examples upfront
A regression model outputs a continuous value, trained on examples where the correct numerical answer was already known in advance. A classification model outputs a category chosen from a fixed set of possibilities, trained the same way, on examples where the correct category was already known. What both share, and what actually makes them both supervised, is that the training data has to already carry the right answer attached before training even starts, which is a genuinely different setup from unsupervised learning, where a model is given no correct answers at all and has to find structure in the data on its own.
The model can only be as good as the labelled data it was corrected against
A supervised model's whole training process is a repeated cycle of predicting, comparing its prediction against the known correct label, and adjusting itself to reduce that gap, which means its eventual accuracy on new, unseen data depends entirely on how good, plentiful and representative its labelled training data actually was. A labelled dataset that's biased or unrepresentative of the population the model will actually be used on doesn't just produce a slightly worse model, it produces a confidently wrong one, since nothing in the training process itself can reveal a systematic gap the training data never showed it.
What we're still unsure about
That supervised learning needs labelled training data and that regression and classification are its two main task types are well established, uncontroversial facts of machine learning. What's more genuinely an unresolved, actively researched problem is guaranteeing that a supervised model trained on one labelled dataset actually generalises fairly to populations that were underrepresented, or entirely absent, in that training data, a model can score extremely well on standard accuracy measures while still making systematically worse predictions for specific groups, and there's no single agreed method yet for reliably catching that gap before a model is actually deployed.
This sits inside Supervised Learning (Regression, Classification), one of eight topics in AI/ML, one of seven domains in Computer Science, one of seventeen subjects the app can quiz you on.