The short answer
- AI detectors estimate how likely a text is to be machine-written by comparing it with patterns learned from human and AI writing. They never see who wrote it.
- Four methods: statistical scores (perplexity and burstiness), zero-shot tests that probe a language model (DetectGPT, Binoculars), trained classifiers (Turnitin, GPTZero, Pangram, Copyleaks) and watermarks hidden at generation time (Google’s SynthID).
- Today’s school tools are classifiers. Turnitin uses a transformer model that scores each sentence from 0 to 1, needs at least 300 words of prose and hides scores between 1% and 19%.
- Yes, AI detectors use AI: the classifiers are themselves deep-learning models.
- Every method produces a probability, not proof. OpenAI withdrew its own classifier in 2023, citing its low rate of accuracy.
AI detectors work by estimating how likely a piece of text is to have been written by a language model. They do not trace where the text came from. Instead they compare its statistical patterns, such as how predictable each word is and how uniform the sentences are, with patterns learned from large collections of human and AI writing, and they output a probability or a percentage.
There are four main ways to do that. This page explains each in plain terms, shows what school tools like Turnitin and GPTZero actually run today, and explains why the same design that catches raw chatbot output also flags some honest students. Whether detectors are accurate enough to act on is a separate question, answered with the evidence in our review do AI detectors actually work?.
The four ways to detect AI text
| Method | How it works | Examples | Main weakness |
|---|---|---|---|
| Statistical scores | Measure perplexity (predictability) and burstiness (variation) using a language model | GLTR (2019); early GPTZero (2023) | Simple, clear human writing scores as “predictable” |
| Zero-shot tests | Probe how a language model reacts to the text, no training data needed | DetectGPT (2023); Binoculars (2024) | Needs access to a similar model; weaker on edited text |
| Trained classifiers | A deep-learning model trained on labeled human and AI text outputs a score | Turnitin; GPTZero today; Pangram; Copyleaks | Only as good as its training data; can’t explain its reasons |
| Watermarking | The AI model hides a statistical signal while generating text | Google DeepMind SynthID Text | Works only for models that apply it; rewriting weakens it |
Table: the four families of AI text detection. Sources: vendor documentation and the research papers linked below.
The first three work after the fact on any text. Watermarking is different: it requires the company that runs the chatbot to cooperate in advance.
What are perplexity and burstiness?
They are the two statistics early detectors used to call text “machine-like.” GPTZero’s own explainer from March 2023 defines perplexity as how likely an AI model would have chosen the exact same set of words, and burstiness as how much writing patterns and perplexity vary across a document.
A language model picks probable next words, so its writing tends to be predictable (low perplexity) and evenly predictable from sentence to sentence (low burstiness). People write more unevenly: a plain sentence, then an odd word choice, then a long tangent.
An illustration (our example, not a detector’s output): after “The results of the experiment were,” a model expects words like “significant” or “inconclusive.” A student who writes “The results of the experiment were a mess, honestly” has picked a far less likely ending, which raises perplexity.
The 2019 research tool GLTR, built by Gehrmann, Strobelt and Rush, colored each word by how highly a model ranked it. In the authors’ human study, it raised people’s ability to spot fake text from 54% to 72% without any training.
The catch is obvious once stated: careful, plain writing is also predictable. That is the root of most false positives.
How do zero-shot detectors work?
They ask a language model how “comfortable” it is with the text, without training a separate detector. Researchers built these methods because classifiers need large labeled datasets and go stale when new chatbots appear.
- DetectGPT (Mitchell and colleagues, Stanford, 2023) makes many small random edits to a passage and checks how the model’s probability score changes. Machine-written text tends to sit at a peak: almost any edit makes it less likely. In the paper, this raised detection performance (AUROC) from 0.81 to 0.95 on one fake-news test set.
- Binoculars (Hans and colleagues, 2024) compares how two closely related language models see the same text. The authors report catching over 90% of ChatGPT samples at a false-positive rate of 0.01% in their tests.
These are lab results on specific datasets. Zero-shot methods matter mostly because they show the signal is real for unedited AI text; school products are built differently.
How do trained classifiers like Turnitin work?
A classifier is a deep-learning model shown huge numbers of labeled examples until it learns to tell the two kinds apart. This is what schools actually use.
Turnitin. According to Turnitin’s AI writing detection FAQ (updated October 2026), its model uses the transformer architecture and is trained on its in-house corpus of academic writing. It is not explicitly programmed to look for burstiness or perplexity; it learns patterns from data. The details:
| Feature | Turnitin’s stated rule |
|---|---|
| Scoring | Each sentence gets a score from 0 to 1, read in context; the document % is the share of sentences above a threshold |
| Minimum length | 300 words of prose (maximum 30,000) |
| What’s ignored | Lists, headers, references, tables, formulas and code |
| Low scores | 1% to 19% shown only as an asterisk, with no highlights |
| Languages | English, Spanish, Japanese, Arabic, Dutch, German, Turkish |
| Paraphrase and “humanizer” detection | English only; added July 2024 and August 2025 |
| Who sees it | Instructors and admins only; admins can turn it off |
Table: how Turnitin's AI writing indicator works, per Turnitin's own FAQ (October 2026) and Plagiarism Today (August 2025).
Turnitin also notes that for short documents of only a few hundred words the prediction is mostly “all or nothing,” because there is only one segment to analyze.
GPTZero started with perplexity and burstiness, but its technology page now describes an end-to-end deep-learning model that classifies text sentence by sentence. Pangram trains a classifier on about a million human documents, each paired with an AI-written “mirror” counterpart, then retrains on the human texts it wrongly flags. Copyleaks says its model learns human writing patterns and flags deviations from them.
So, do AI detectors use AI? The ones in schools do. They are neural networks judging text produced by other neural networks.
What is AI watermarking?
A watermark is a hidden statistical signal added while the chatbot writes, so the text can be checked later. In the method by Kirchenbauer and colleagues (2023), before each word is chosen the model marks a random set of “green” tokens and nudges its choices toward them. A detector that knows the scheme counts green tokens and runs a statistical test; it does not need the model itself.
Google DeepMind’s SynthID Text, described in Nature on October 23, 2024, adjusts word probabilities at the sampling step. In a live test covering nearly 20 million Gemini responses, users rated watermarked and unwatermarked answers about the same, and Google released the code the same day. DeepMind says the signal survives light edits but weakens with thorough rewriting or translation.
Why hasn’t watermarking solved the problem? A watermark only works for models that use it, and students can use many. OpenAI said in August 2024, as reported by TechCrunch, that it had an accurate text-watermarking method but was not releasing it: it called the method trivial for bad actors to circumvent and worried it could stigmatize AI use by non-native English speakers.
What do the vendors themselves say about accuracy?
Vendors publish strong numbers, and every one of them also says the score is not proof.
| Vendor | Claim (self-reported) | Their own caution |
|---|---|---|
| Turnitin | Under 1% false positives for documents with over 20% AI writing; tested on 700,000+ pre-ChatGPT papers | Score “should not be used as the sole basis for action”; a 50% score could mean up to 65% AI text |
| GPTZero | 96.5% accuracy on mixed documents; false positives no more than 1% (1.1% on ESL essays) | Its FAQ says no AI detector is 100% accurate |
| Copyleaks | 99% accuracy; 0.03% false-positive rate | Warns that Grammarly-style AI features can be flagged |
| OpenAI (2023, withdrawn) | Caught 26% of AI text; flagged 9% of human text | Withdrawn in July 2023 for “its low rate of accuracy” |
Table: vendor accuracy claims. Sources: Turnitin FAQ; gptzero.me/technology; copyleaks.com; TechCrunch (Jan. 31, 2023); Decrypt (July 2023).
Independent studies are much less flattering, especially for edited text and for writers working in a second language. We cover those numbers in what six studies found about AI detector false positives rather than repeating them here.
Why do AI detectors flag human writing?
Because every method measures how ordinary the writing is, not who wrote it. Predictable vocabulary and even sentence structure are exactly what a model learns to associate with AI, and many humans write that way on purpose: students following a five-paragraph template, writers in their second language, and anyone polishing text with grammar tools. Turnitin’s FAQ lists little structural variation, repetitive text and paraphrasing without new ideas as common triggers, and says Grammarly’s generative features “will likely be flagged.”
Symptom, cause and what helps:
| What happened | Why the detector reacted | What helps |
|---|---|---|
| A formulaic essay scored high | Template structure is low-variation by design | Drafts and notes that show the process |
| Short answer flagged entirely | Under a few hundred words, scores are close to all-or-nothing | Don’t treat short-text scores as meaningful |
| Grammar app used, then flagged | Generative rewriting changes word patterns | Disclose tools; keep the pre-edit draft |
| Second-language writer flagged | Simpler vocabulary reads as predictable | Version history and a conversation about the work |
Table: common reasons for AI detector false positives, based on Turnitin's FAQ and the research on non-native writers.
What this means for students and teachers
Treat a detector score as a reason to look closer, never as a verdict. That is what the vendors say, and it follows from how the tools work: they output probabilities based on surface patterns.
For students, the best protection is a visible process: version history in Google Docs or Word, notes, outlines and the ability to explain your argument out loud. Our guide to using AI for homework without cheating covers disclosure, and how to cite ChatGPT shows the formats. If you want to know whether your school or college runs one of these tools, see which AI detectors colleges use.
For teachers, detecting AI work honestly explains what signals are more reliable than a score, and our classroom AI policy template helps set the rules before anyone is flagged. For a one-line definition, see the glossary entry for AI detector.
An honest caveat: detection technology changes every few months. Turnitin updated its English model architecture in October 2026, and vendor accuracy figures are self-reported. The mechanisms on this page change more slowly than the numbers.