A label is the correct answer written beside a data example, the thing the model is meant to learn. Supervised learning runs on this. The model sees the input, makes a prediction, compares it to the label and adjusts itself to shrink the difference.
The form a label takes depends on the task. In classification it is a category name, such as whether an email is spam. In regression it is a number, such as the sale price of a flat. In object detection it is a bounding box plus the name of what sits inside it. In sentiment analysis it is the attitude the text carries. In ranking tasks a label can be which of two options a person preferred.
Producing labels is usually the most expensive part of the job. Specialist domains need doctors, lawyers or subject experts doing the work. That cost is why methods working from few labels, and approaches that derive labels from existing data, spread so quickly.
A concrete case: a bank builds a fraud detection model. The label comes from whether a transaction was later disputed. That label arrives late, since a dispute is sometimes recorded two months afterwards. If the delay is ignored during training, the model is being taught with information it will not have in production.
When a label is wrong the model learns the error. High accuracy on a badly labelled set means nothing more than imitating the mistakes well.