Zero-shot: asking a model to do something it has never been shown
Zero-shot means giving a model a task with no examples of what a good answer looks like. You say "translate this sentence into Swahili" and it does, without you first showing it three translated pairs. Few-shot means you include a handful of examples; one-shot means one. None of these is training. They are ways of prompting an existing model. Humanize 360 does not prompt a model at all, but the term matters because it also names a family of AI detection methods, and the two meanings get mixed up.
The prompting sense is easy to picture. Zero-shot: "Classify this review as positive or negative: The battery died in a week." Few-shot: the same request preceded by two labelled reviews. The examples do not teach the model anything permanent; they nudge it toward the format and the boundary you want. For simple tasks zero-shot works fine. For unusual formats or fine distinctions, examples help a lot.
The second meaning: zero-shot detection
In AI detection, a zero-shot method is one that does not need a training set of labelled human and machine texts. Instead it measures properties of the text directly, usually with a reference language model: how probable each word was, how the probability changes if you perturb the text, how even the rhythm is. Perplexity-based detection is the best known example. The alternative, a trained classifier, learns from labelled samples and can be sharper on the kind of text it saw and weaker on everything else.
Why the distinction matters to you
- Zero-shot detectors generalize to new models better, because they were never tuned to one model's quirks.
- Trained classifiers can be more accurate on familiar genres and more confidently wrong on unfamiliar ones.
- Most commercial detectors blend both and do not say how much of each.
- Neither kind can verify its verdict; both are estimates from texture.
Where Humanize 360 sits
Our detector is closer to the zero-shot family. It computes eleven named signals directly from the text (sentence-length variance, word predictability, stock-phrase density, opener repetition, tell-word ratio, punctuation patterns, paragraph uniformity, contraction rate, vocabulary spread, transition-word overuse and structural sameness) and combines them into a Human Score. There is no black-box classifier trained on a secret sample, which is why every signal can be shown to you and argued with.
A caution about the word
Vendors like "zero-shot" because it sounds effortless. In prompting it just means "no examples". In detection it just means "no labelled training set". Neither implies accuracy, and a zero-shot detector still depends on which reference model it uses to judge predictability. When you read the word in marketing copy, ask which sense is meant and what it is being contrasted with.
Common questions
Is zero-shot better than few-shot?
For simple, common tasks they perform similarly. For unusual formats or subtle categories, a few examples usually improve results. Try zero-shot first and add examples only where the output misses.
What is zero-shot AI detection?
Detection that measures the text directly, typically with a reference model, instead of learning from a labelled set of human and machine samples. Perplexity and burstiness measures are the common examples.
Is Humanize 360's detector trained on ChatGPT output?
No. It measures eleven signals computed from the text itself, with published thresholds. That makes it explainable, and it means it is not tuned to any single model's quirks.
Does zero-shot mean the model has never seen the topic?
No. It means no examples were provided in the prompt. The model may have seen plenty about the topic during training; that is a separate matter.