Zero-Shot

If you tell a computer program: “Translate this sentence into Spanish,” you would normally expect that someone built it specifically for that purpose. With modern AI systems, this is no longer necessary. You can simply describe a task to them in plain language, and they will try to solve it. Zero-shot is the name for this case: zero examples, just one instruction. The program has never previously been shown what a correct solution to this specific task looks like. The name comes from the English word “shot,” which here stands for an example provided along with the request.

Why zero examples save so much work

In the past, machine learning was a laborious affair. Anyone who wanted a spam filter had to manually label tens of thousands of emails as “spam” or “not spam.” Only with this labeled data could a system be trained. For every new task, the data collection had to start all over again.

Zero-shot reverses this relationship. A large language model — an AI system trained on huge amounts of text from the internet — can take on many such tasks without any preparation. Instead of weeks of data work, all that’s needed is a well-formulated sentence. This is the main reason why AI has appeared in so many products since around 2022: the barrier to entry has dropped from “research project” to “one afternoon.”

The flip side is important, though. A system trained specifically for a task is usually more accurate than a zero-shot solution. For medical findings or legal documents, “pretty good” is not good enough. Zero-shot is therefore often the quick first step, not the final solution.

How the model knows the task anyway

The trick lies in the pretraining. The model didn’t learn to solve a particular task, but rather to continue text. In doing so, it processed billions of sentences: translations, summaries, product reviews, recipes, program code. The patterns of these tasks were already embedded in these texts, without anyone having marked them as exercises.

You can compare it to a student who has read a great deal. He never practiced writing book reviews specifically. But because he has read hundreds of them, he can write one right away that roughly fits the bill. This exact “roughly fits, without practice” is the zero-shot ability.

Alongside this are two related terms. With one-shot, you provide a single example; with few-shot, a few, usually two to five. These examples appear only in the request itself — the model itself is not modified. Often the success rate increases noticeably as a result, especially when a particular output format is required.

Zero-shot in chatbots and in technical reports

You most often encounter zero-shot without noticing it. Every question to a chatbot is essentially a zero-shot request. The translation function in your browser, subject-line suggestions in your email program, or automatic image descriptions in a photo app also often work this way. There is no special model behind it, just a general model with a fixed instruction.

In technical articles and company announcements, these terms appear as metrics. If it says “82 percent zero-shot accuracy,” that means: the model was tested on a task it was not trained for. Such figures are considered an indicator of how well a model handles novelty. Comparisons between providers are common here — and one should pay attention to whether zero-shot or actually few-shot was measured, since the numbers differ substantially.

A common misconception: zero-shot does not mean the model has learned nothing. It has learned a great deal, just nothing specifically for this task. Nor does it mean the answer is correct. A model can answer a completely unfamiliar task confidently and wrongly.