label
/ LAY-bul /
A label is the right answer attached to an example — the thing you want the machine to learn to predict. If you show a program thousands of photos and tell it which ones contain a cat, the word "cat" or "no cat" pinned to each photo is its label. It is the back of the flashcard: the input is the question, the label is the answer the machine checks itself against while it studies.
Labels come in different shapes. They can be a category, like "spam" or "not spam" (used for classification), or a number, like tomorrow's temperature (used for regression). Where they come from matters enormously. Most labels are created by people, one example at a time — a slow, costly job called data annotation. Some labels arrive for free from the world itself: whether a customer actually clicked the link, whether the patient recovered. Those natural outcomes are often the most trustworthy labels of all.
The honest catch is that a model is only as good as its labels. If the labels are wrong, inconsistent, or biased, the machine will faithfully learn those mistakes — garbage in, garbage out. Two people may even disagree on the correct label for the same example (is that review positive or just sarcastic?), and that built-in disagreement sets a ceiling on how accurate any model can ever be. Many tasks have no labels at all, which is exactly the situation that unsupervised learning is designed for.
A team wants to detect tumours in X-rays. Radiologists go through 50,000 scans and mark each one "tumour" or "clear." Those marks are the labels. When two radiologists disagree on a borderline scan, that single ambiguous label quietly limits how good the final model can be.
Labels are the answer key — and human disagreement over them caps a model's accuracy.
"Label" usually implies supervised learning, where every example comes with its answer. The most expensive part of many AI projects isn't the computing — it's the painstaking human work of producing accurate labels.