Zero-shot learning is asking a model to handle a task it was never explicitly trained or shown examples for, relying only on a plain-language instruction. You describe what you want, such as classify this review as positive or negative, and the model does it from general knowledge.

It matters because it sets the floor for how useful a model is out of the box. Strong foundation models can handle a surprising range of tasks with no examples, which makes prototyping fast and cheap. Accuracy is usually lower than with examples, so it is a starting point rather than a finish line for high-stakes work.

At arosplatforms we use zero-shot prompting to quickly gauge whether a model can do a job at all, then add a few examples or grounding when more reliability is needed. It is the cheapest experiment in the toolkit and a fast way to size up a problem.