What AI detection is
AI detection estimates how likely it is that a passage of text was produced by a large language model rather than written by a person. It does not read meaning, check facts, or search a database of known AI outputs. There is no registry of everything ChatGPT has ever written to compare against.
Instead, detection works statistically. Language models generate text by repeatedly predicting the most probable next word given everything written so far. That process is extremely good at producing fluent prose — and it leaves behind measurable regularities that human writing tends not to have. Detection looks for those regularities.
Why results are probabilistic, not binary
This is the single most important thing to understand, and the point most often lost in practice.
Human and machine writing are not two cleanly separated populations. They overlap. Some people write in a highly regular, structured way that statistically resembles model output — technical documentation, formal academic prose, and writing by people composing in a second language are common examples. Meanwhile, AI text that a person has revised takes on the irregularity of human editing.
What follows from this: a detection score is evidence, not proof. It should inform a human decision, never replace one. We say this plainly because the alternative — treating a percentage as a verdict — leads to real harm, particularly in education.
The patterns language models leave behind
Two broadly documented properties of machine-generated text do most of the work in detection. Both are concepts from the published research literature, not proprietary ideas.
Predictability
Because a model is optimised to choose likely continuations, its output tends to sit closer to the “expected” word at each step than human writing does. Researchers refer to this using the concept of perplexity — loosely, how surprised a language model is by a piece of text. Machine-generated passages typically produce lower surprise. Human writing wanders more: it reaches for an unusual word, abandons a construction mid-thought, repeats itself imperfectly.
Uniformity of rhythm
The second property is often called burstiness — the degree to which sentence length and complexity vary across a passage. Human writing is bursty: a long, winding sentence followed by three words. Model output tends toward a steadier cadence, with sentences of similar length and comparable structural complexity. Neither pattern is decisive alone, but together they are informative.
A useful analogy: handwriting analysis does not read what a note says. It looks at pressure, slant and spacing. Detection works the same way — it examines the texture of the writing rather than its content.
How ZeroGPT analyses your text
When you submit text, it moves through a sequence of stages rather than a single pass. Each stage narrows the question.
Preparation
The text is normalised and split into sentences. Formatting artefacts — stray line breaks, pasted markup, inconsistent spacing — are cleaned so they do not distort the analysis. Very short inputs are flagged here, because they cannot support a reliable result.
Sentence-level analysis
Each sentence is assessed on its own. This is why results are shown as a highlighted breakdown rather than one number: mixed documents are the normal case, and a single figure would hide where the signal actually is.
Passage-level consistency
Sentences are then examined together. Does the vocabulary range stay uniform? Does sentence rhythm vary the way human drafting usually produces? Abrupt shifts in style mid-document are themselves informative — they often mark the boundary between written and generated material.
Context weighting
Signals are weighted by how much evidence is actually available. A long, coherent passage yields a confident reading. A three-sentence fragment does not, and the result reflects that rather than overstating certainty.
Scoring
The weighted signals are combined into an overall likelihood, alongside the per-sentence breakdown that lets you see how that figure was reached.
The sentence-level view matters more than it first appears. A document that is 30% machine-generated is a genuinely different situation from one that is uniformly 30% “suspicious” throughout — and only a breakdown distinguishes them.
Reading the confidence score
The score answers one question: how strongly does this text exhibit the statistical patterns associated with machine generation? It is not a percentage of the document written by AI, and it is not a probability that a specific person cheated.
| Result | Reasonable reading |
|---|---|
| High | Strong machine-like patterns throughout. Worth reviewing — alongside other context. |
| Mixed | Patterns vary across the document. Use the sentence breakdown; a uniform judgement is not appropriate here. |
| Low | Few machine-like patterns. Note that this does not rule out AI text that was substantially rewritten. |
Known limitations
We publish these because a methodology page that omits them is not a methodology page.
- Short texts: Below roughly a paragraph, there is not enough signal for a dependable reading. Treat short-text results as weak evidence.
- Heavily edited AI text: Meaningful human rewriting removes the statistical fingerprint. No current detection method is immune to this.
- Non-native English writing: Simpler structures and more predictable vocabulary can resemble machine output. This is a well-documented source of false positives and a specific reason not to apply results mechanically.
- Highly formulaic writing: Legal boilerplate, technical specifications and formal academic prose are naturally regular, which can raise scores for genuinely human text.
- Very new or unusual models: Output from a model family the system has had little exposure to may be assessed less accurately until coverage catches up.
- Translated text: Machine translation of human writing introduces its own regularities and can shift results in either direction.
For educators specifically: please do not treat a score as the basis for an academic-integrity finding on its own. Use it as one input — with draft history, prior work, and a conversation with the student. False positives are a real phenomenon and they fall hardest on students who write plainly or in a second language.
Model coverage and continuous improvement
Detection is adversarial by nature. Each new generation of language models writes with more human-like variation than the last, which means a detector left unchanged degrades over time. Ours is evaluated and updated on an ongoing basis as new model families become widely used — including the current major systems from OpenAI, Anthropic, Google, DeepSeek and Meta.
If you use detection at scale, our API exposes the same analysis programmatically, and higher-volume plans are available for institutional use.
Privacy
Text submitted for detection is analysed and then discarded. We do not retain submitted content to train our models, and we do not share it with third parties. This matters practically as well as ethically: educators handling student work and publishers handling unreleased drafts need to know where that text ends up. Full details are in our privacy policy.
Try it on your own text
The clearest way to understand a detection score is to run text you already know the origin of — something you wrote, and something you generated — and compare the sentence-level breakdowns.