Zero-shot learning is a model performing a task without having seen a single example of it. No example goes into the prompt, only a plain description of what is wanted. The model reads the description, works out the task and produces output directly.
What makes this possible is the breadth of pre-training. The model has already seen countless task shapes across an enormous body of text: summarising, classifying, translating, filling in tables. Describe a new task and it relates that description to patterns it has seen. On the vision side, a model that has never seen a class can still recognise it when the class is described in words.
The difference from few-shot is the example count. Few-shot puts several examples in the prompt, zero-shot puts none. That means shorter prompts and lower cost, with less control over the shape of the output.
A concrete case: a publisher wants news headlines sorted into subject categories. The prompt carries only the category list and an instruction to assign each headline to one of them. The model produces sensible results without seeing a single labelled example. Accuracy drops once the categories involve in-house jargon, and a few examples become necessary.
The weak spot is format consistency. The model sometimes strays outside the requested label set, so output gets constrained by a schema or validated afterwards.