Zero-Shot Learning — Definition, the Two Senses of the Term, and What It Removes From a Chatbot Project (2026)

Quick answer: Zero-shot means the system is asked to recognize something it was given no labeled examples of, working instead from a description. Ask a large language model "which of these five categories does this message belong to, and here is what each category means," and you have zero-shot classification. Nobody wrote training phrases. That is genuinely useful, and it is also why the label text you write becomes the most important artifact in the project, because it is now the only part of the classifier you can edit. The work did not vanish. It changed shape, from collecting examples to writing definitions, and writing definitions is harder than most teams expect on the day they discover the boundary between "cancel my order" and "change my order" was never stated anywhere.

Two different things get called zero-shot

The term has two lineages, and they answer different questions. Mixing them is why the same phrase can sound rigorous in a paper and vague on a pricing page.

The classical sense comes out of machine-learning research from the late 2000s. The problem it names is a classifier that has to predict categories missing from its training set entirely. Palatucci, Pomerleau, Hinton and Mitchell put it plainly in their 2009 NeurIPS paper: "We consider the problem of zero-shot learning, where the goal is to learn a classifier f: X → Y that must predict novel values of Y that were omitted from the training set." Their route through the problem was a knowledge base of semantic properties, so that a category the model had never seen could still be described in terms it had. The case study was reading words out of fMRI scans of people thinking, including words with no training scans behind them.

The in-context sense is the one that reached chatbot platforms, and it arrived with GPT-3. Here zero-shot describes what is absent from the prompt rather than what is absent from the training set. The model gets a description of the task and no worked examples, and nothing about the model changes. The 2020 paper's abstract states the setup directly: "For all tasks, GPT-3 is applied without any gradient updates or fine-tuning, with tasks and few-shot demonstrations specified purely via text interaction with the model." The same abstract names the cost it was trying to avoid, and the number is the reason anyone cared: conventional fine-tuning "still requires task-specific fine-tuning datasets of thousands or tens of thousands of examples."

What both senses share is the substitution at the center. A category that would normally be demonstrated is instead described. Everything practical about zero-shot in a chatbot follows from that swap, because a description is a piece of writing, and pieces of writing are the sort of thing that can be vague.

Two neighbors are worth separating while we are here. Few-shot puts a handful of examples in the prompt, which is not training either, since the model's weights are untouched and the examples are gone the moment the request ends. Fine-tuning does change the weights, and it is the only one of the three that produces a new artifact you have to version and maintain. The ladder from zero-shot up through few-shot to a trained classifier, and the question of when a bot has earned the next rung, is the subject of our companion guide; this entry stops at what the term means and what it moves.

What "no training data" is actually claiming

Read the vendor sentence carefully and it turns out to be scoped. No training data is required from you. Enormous quantities went into the model, which is the entire reason it can classify anything at all, and that data is not visible, not editable, and not itemized on your invoice.

This matters for three ordinary reasons.

The capability you are getting is the model's, not the platform's. Swap the model underneath and the accuracy moves, which is why a platform that lets you choose the model is offering something different from one that does not. Our review of Botpress records native routing across OpenAI, Anthropic, Groq and Hugging Face as first-class providers, with different tasks routable to different providers, so intent classification can run on a different model from generation. That is a genuine lever. On most marketing-automation platforms the model is vendor-managed, and the lever does not exist.

Your domain is probably underrepresented in whatever the model absorbed. Generic customer-service language is everywhere on the public internet. Your product names, your internal shorthand, the two plan tiers your customers keep confusing are not. Zero-shot is strongest exactly where your business is least distinctive, which is the opposite of where you needed the help.

And nothing about the arrangement is stable by default. The examples live in a model the vendor can update. A classification that was right in March can be right in a different way in September, which is a maintenance property rather than a defect, and it is the reason the QA protocol asks for a fixed regression set instead of a launch-day spot check.

The label is the training data now

Here is the practical inversion, and it is the single most useful thing to carry out of this entry.

In a conventional intent recognition setup, you supply examples and the platform derives the boundary between categories from them. Fifteen phrasings of "where is my order" teach the classifier what that intent looks like, and the boundary is learned. In a zero-shot setup, there are no examples, so the boundary has to be stated. Whatever you typed into the intent name and description is the whole specification.

Which means a label like cancel_order is not a specification. It is a guess that the model will draw the line where you would have. Ask it to choose between cancel_order, change_order and return_order on the message "I don't want this anymore, can you sort it out" and it will pick one, confidently, and there is no fact of the matter in your configuration that says which is correct. The fix is not more machinery. It is writing the labels so each one states its own edge: cancel applies before dispatch, change applies before dispatch and keeps the order alive, return applies after delivery.

That rewrite is prompt work, and it behaves like prompt work, which is to say it is fast to iterate and easy to get wrong in ways that only show up in transcripts. Our prompt engineering entry covers the general discipline and the chatbot-specific guide covers the applied version. The habit worth forming immediately is to write labels for the confusable pairs first, since a category with no near neighbor rarely causes trouble and two adjacent categories cause most of it.

Evidence that the wording carries this much weight is not hard to find in the research either. Kojima and colleagues reported in 2022 that inserting the single phrase "Let's think step by step" before the answer raised accuracy on one arithmetic benchmark "from 17.7% to 78.7%" and on another "from 10.4% to 40.7%" with the same model and no examples added. That result is about reasoning tasks rather than intent labels, so read it as an illustration of sensitivity rather than a transferable number. The illustration still lands: in a zero-shot system, the text is the configuration.

Where zero-shot degrades, and what each failure is telling you

Zero-shot fails in patterns, and the pattern identifies the fix. The table below is our editorial reading of the failure modes rather than a published taxonomy, assembled from the classification behavior we work through when evaluating these platforms.

The third row is the one that costs real money, because a zero-shot classifier with no out-of-scope option is a machine for producing confident wrong routing. The fallback intent entry covers the escape hatch and our fallback-rate guide covers what a healthy rate looks like. A bot that never falls back is not accurate. It is silent about being wrong.

The fourth row deserves its own note for anyone selling outside one language. Zero-shot is often marketed as multilingual out of the box, and the honest version is that it is multilingual to the extent the underlying model is, which varies by language and is rarely published per language. Our multilingual guide works through what that means for a real deployment, and language detection covers the step that has to happen before classification does.

What our own testing shows about zero-shot accuracy

Across the platforms we have reviewed hands-on, the intent-accuracy figures we measured are zero-shot figures. In none of those sessions did we write training utterances first. Where we supplied anything, we supplied content, meaning help articles and PDFs for the retrieval layer, and the classification itself ran on descriptions rather than demonstrations. The band those tests produced on twenty-query sets sits roughly between the high seventies and high eighties depending on product and language, and the per-language spread inside a single product is the part worth staring at, because it is the clearest available evidence that the capability is inherited from the model rather than configured by you.

The specifics, with their evidence depth attached, since it differs: our Chatbase review measured 89 percent English, 85 Spanish, 83 Brazilian Portuguese and 79 Hindi on a Standard-tier account, the strongest set in that batch. Manychat measured 89 English, 84 Spanish and 82 Brazilian Portuguese on a Pro tier, and its builder is the clearest illustration of the shift this entry describes, because the trigger sidebar offers to convert a keyword list into an intention directly rather than asking for training phrases. Tidio measured 87 percent on English and Polish and 80 on Spanish and Portuguese, and its own review records that the vendor does not publish a certified per-language list, which is the transparency gap this section is about. Botpress was tested on Plus with Claude 3.5 Sonnet configured as the provider, and its numbers move with that choice by design. Two caveats belong here rather than in a footnote: the Intercom figures in that review are anchored editorial estimates with paid-tier hands-on validation scheduled rather than complete, and our SendPulse review has no measured per-language accuracy at all, with that pass still queued. Treat those two differently from the four that were measured. Every one of these reviews carries affiliate links, and our methodology sets out what that does and does not change about a score.

Four questions for a vendor page that says "no training required"

1. Can I edit the description of each category, or only its name? If the platform gives you a name field and nothing else, your entire specification is one or two words, and the confusable pairs will stay confusable.
2. Do I see a confidence number, or only the winning category? A classifier that reports only its answer cannot be given a threshold, and without a threshold there is no principled route to a clarifying question.
3. Is out-of-scope a real category or an afterthought? Ask what happens to a message that fits nothing. "It picks the closest one" is a description of a problem.
4. Which model classifies, and can I pin it? This decides whether your accuracy is a number you own or a number that moves when the vendor upgrades. Platforms with BYOLLM support answer it one way and vendor-managed platforms answer it another, and neither answer is wrong, but only one of them is under your control.

The line to carry out of all four: zero-shot removes the chore of collecting examples and hands you the harder job of writing definitions that hold up against messages nobody anticipated. That is usually a good trade at the start of a project, when you have no examples to collect anyway. Whether it stays a good trade at month three is a separate question, and it is the one our companion guide exists to answer.

Related terms

• Intent recognition — the task zero-shot classification most often performs inside a chatbot, and the entry that covers the classifier alternative.
• Chatbot training — the umbrella word platforms use, which now describes four different mechanisms, only one of which involves examples you supply.
• Large language model — where the zero-shot capability actually comes from, and why it moves when the model does.
• Fine-tuning — the far end of the same axis, and the only option on it that changes the model rather than the prompt.
• Utterance — the unit of training data that zero-shot lets you skip, and the unit you will still need for testing.

FAQ

What is zero-shot learning in simple terms?

Asking a model to recognize something you never showed it examples of, by describing the thing instead. Tell it that a message belongs in one of five buckets and explain what each bucket means, and it will sort messages into buckets without ever seeing a sorted message. The trade is that your description is now doing the job examples used to do, so a vague description produces vague sorting.

Does zero-shot learning mean no training data at all?

No, and this is the sentence worth being precise about. It means no task-specific labeled data from you. The model absorbed a very large amount of text during pretraining, which is the only reason it can classify anything, and none of that is visible or editable on your side. The honest reading of "no training required" is that the vendor has removed a chore from your project, not that the system learned your categories from nothing.

What is the difference between zero-shot and few-shot?

The number of examples inside the prompt: none versus a handful. Neither changes the model, and neither persists, because both are instructions sent with a request rather than an update to anything. Few-shot usually raises accuracy on the categories the examples illustrate, costs a little more per call because the prompt is longer, and is the natural first move when a specific pair of categories keeps getting confused. Our companion guide covers when to make that move.

Can I build a chatbot with zero-shot intent classification and no training phrases?

Yes, and on several of the platforms we review that is the default path rather than an advanced option. What you cannot skip is testing. Zero-shot removes the training set and leaves the test set exactly where it was, because the only way to know whether your category descriptions hold is to run real messages through them and read the results. Twenty to thirty real questions with known-correct answers, re-run whenever you edit a label, is the minimum version of that.

Is zero-shot classification accurate enough for customer support?

For routing and first-line answering, our measured figures put well-configured products in the high seventies to high eighties on twenty-query sets, which is workable when there is a confidence floor and a human handoff beneath it, and dangerous when there is not. The number that decides the question is not the accuracy figure. It is what happens on the remaining share, because a bot that routes fifteen percent of contacts to the wrong place silently is worse than one that routes them to a person.

Is zero-shot the same as retrieval-augmented generation?

No, though they often ship together and get described with the same marketing sentence. Zero-shot is about how a category or task is specified, with a description standing in for examples. Retrieval-augmented generation is about where the facts in an answer come from, with your documents fetched at request time. A typical 2026 support bot does both: zero-shot to work out what is being asked, retrieval to answer it. They fail differently, so it is worth knowing which one produced a bad transcript before you go fixing anything.

Sources

• Palatucci, Mark, Pomerleau, Dean, Hinton, Geoffrey E. and Mitchell, Tom M. Zero-shot Learning with Semantic Output Codes. Advances in Neural Information Processing Systems 22 (NIPS 2009), pp. 1410-1418. proceedings.neurips.cc/paper/2009/hash/1543843a4723ed2ab08e18053ae6dc5b-Abstract.html (abstract fetched and quoted verbatim 30 July 2026). Source of the classical formulation quoted above.
• Brown, Tom B., Mann, Benjamin, Ryder, Nick, et al. Language Models are Few-Shot Learners. arXiv:2005.14165, submitted 28 May 2020, v4 of 22 July 2020. arxiv.org/abs/2005.14165 (abstract fetched and quoted verbatim 30 July 2026). Source of the in-context formulation and of the thousands-to-tens-of-thousands figure for conventional fine-tuning datasets.
• Kojima, Takeshi, Gu, Shixiang Shane, Reid, Machel, Matsuo, Yutaka and Iwasawa, Yusuke. Large Language Models are Zero-Shot Reasoners. NeurIPS 2022; arXiv:2205.11916, v4 of 29 January 2023. arxiv.org/abs/2205.11916 (abstract fetched and quoted verbatim 30 July 2026). Source of the 17.7 to 78.7 percent and 10.4 to 40.7 percent figures, both reported on reasoning benchmarks with the text-davinci-002 model, and of the paper's closing recommendation to examine zero-shot capability "before crafting finetuning datasets or few-shot exemplars."
• Chatbotscape Glossary. Large language model — carries the primary citations for model architecture and behavior inherited by this entry.
• Chatbotscape Glossary. Intent recognition — carries the classifier-versus-LLM comparison and the worked example this entry deliberately does not repeat.
• Chatbotscape platform reviews for Chatbase, Manychat, Tidio, Botpress, Intercom and SendPulse, last updated between 26 and 31 May 2026. Intent-accuracy figures are stated as of each review's own date and differ in evidence depth: the Chatbase, Manychat, Tidio and Botpress figures are hands-on measurements on twenty-query sets, the Intercom figures are anchored editorial estimates with paid-tier validation scheduled, and the SendPulse review carries no measured per-language accuracy, with that pass queued.
• Chatbotscape evaluation methodology. /methodology (continuously updated).