Machine learning from examples that come paired with the right answers.
Supervised learning teaches an AI with examples that carry answer labels: "this photo is a cat", "this email is spam". The AI makes a guess, compares it with the label, and adjusts its numbers by how wrong it was. Sorting photos, filtering spam and reading handwriting are classic uses.
The catch: labeling takes a lot of human work. So it's used alongside unsupervised learning, which finds patterns without labels, and self-supervised learning, which makes its own quiz questions from raw data. Large language models first learn self-supervised, guessing hidden next words, then get polished with supervised learning on answers written by people.
Think of it like this
It's like studying with a workbook that has an answer key. You do a problem, check the key, and fix what you got wrong. Making a good answer key takes effort, which is why good workbooks are precious.
For example
Show an AI thousands of photos labeled "dog" or "cat", and it learns to sort new photos it has never seen.
Common mix-ups
Mix-upIn supervised learning, the AI learns by listening to a teacher's explanations.
ActuallyHere "supervised" doesn't mean someone explains things; it's closer to being checked against the right answers. A label says only "what", never "why", so the AI has to guess the reason from what the examples have in common.
Mix-upLarge language models are built with supervised learning from start to finish.
ActuallyEven after tuning with model answers, they often go through a further reinforcement learning stage using records of people choosing the better of two answers. That stage gives only comparisons — "this one's better" — not a single right answer, so it isn't supervised learning.
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- If some labels are wrong, what will the AI learn?
- Activity
- Each group makes a "workbook" of ten dog and cat picture cards with labels — but one group secretly swaps two labels. Another group reads only the workbook, writes a three-line rule for telling dogs from cats, and applies it to five new cards. Compare how the rule built from the mislabelled workbook turned out. Then time how long labelling ten cards took and estimate it for 1,000 photos — that puts a number on how much human work labels take.