Data
A label is the correct answer attached to a training example, such as the tag 'cat' on a photo, that a supervised model learns to predict.
For a model to learn from examples, someone has to say what each example is. That answer is the label. A photo gets 'cat' or 'dog.' An email gets 'spam' or 'not spam.' A sentence gets 'positive' or 'negative.' A house record gets its sale price. Labels are what turn raw data into teaching material.
Labeling is often done by people, one item at a time, which is slow and costly. Big projects hire thousands of annotators. Labels can also be noisy: two people may disagree about whether a review is sarcastic. Those disagreements become confusion for the model.
Not all learning needs labels. Unsupervised learning finds structure without them, and language model pretraining uses the text itself as the answer. But most business AI, from medical imaging to fraud detection, still rests on carefully labeled data.
A radiologist marks 5,000 chest X-rays as 'normal' or 'shows pneumonia' so a model can learn to flag the second kind.