How Do AI Detectors Work? Perplexity, Classifiers and Watermarks Explained

Paste a paragraph into an AI detector and a number comes back: 78% likely AI. It looks precise. It isn't. The tool hasn't checked who wrote the text, when, or with what. It has measured how much the writing resembles what a language model tends to produce, and turned that into a percentage.

This guide explains how that measurement works, why it fails in predictable ways, and how to read a result without fooling yourself. We build and train an AI detector ourselves, so some of the examples come from our own work.

How do AI detectors work?

An AI detector looks at a piece of text and estimates how likely it is that a language model wrote it. It does this in one of three ways: by scoring how predictable the words are, by running a classifier trained on labelled human and AI examples, or by checking for a hidden watermark. AI checkers, AI content detectors and AI writing detectors are all names for the same thing.

Method 1: perplexity and burstiness

Language models write by picking a likely next word, over and over. So text a model produced tends to be text a model would have predicted. Early detectors used that directly:

• Perplexity is how surprising each word is, given the words before it. Low perplexity means nearly every word was the obvious next word.
• Burstiness is how much sentence length and rhythm vary. People write a long sentence, then a short one. Then a fragment. Models tend to keep a steady 15 to 25 words.

These two ideas are still the clearest way to explain AI detection, but they're no longer the whole of it. GPTZero, which made the terms popular, says it moved to a deep-learning model in autumn 2023 and now treats perplexity and burstiness as one of seven indicators.

Method 2: trained classifiers

Most detectors today are classifiers. You show a model a large set of texts labelled "human" and "AI", and it learns which patterns separate them. Those patterns can be character sequences, word pairs, sentence rhythm, punctuation habits and favourite phrases. No single pattern decides the score. Thousands of small signals add up.

The catch is that a classifier learns "human" from whatever human writing it was shown. We found that out the hard way. On 2 October 2026 we tested our own detector on 140 texts in seven sets. It was clean on modern news. But it wrongly flagged 7 of 36 passages of older literature, and it caught only 1 of 10 AI texts written in Traditional Chinese, because its Chinese training data was almost all Simplified.

Both problems came from the training data, not from the writing. Once we converted Traditional Chinese to Simplified before scoring, the detector caught all 10 of those texts. After the retrain on public-domain books that followed, 9 of the 10 are flagged and the tenth scores 57, which the tool calls Mixed. Its false alarms on older literature by unseen authors fell from 31.6% to 2.5% in Japanese and from 22.4% to 6.0% in French. Any detector carries blind spots like these. Most vendors just don't publish theirs.

Method 3: watermarks

A watermark is a signal added while the text is being written. Google DeepMind's SynthID Text nudges the probabilities of the words Gemini picks, in a pattern a detector can later check for. It ran on Gemini from May 2024 and was made open source in October 2024, with the method published in Nature. In a live test on about 20 million Gemini responses, users didn't notice a difference in quality.

Watermarks have two big limits. They only detect text from a model that added the watermark. And Google DeepMind says SynthID is weaker on text that was rewritten or translated, and on short factual answers where there's little room to vary the wording.

A worked example

Here are two sentences that say the same thing. Guess which one a detector flags.

Version A reads as machine-like to almost any detector. Every word is the expected word, the sentence is a textbook length, and nothing in it is something only its author could know. Version B reads as human: a month, a doctor, a small joke.

Now the uncomfortable part. Version A could have been written by a nurse with 20 years' experience. Version B can be produced by any chatbot told to "add a personal anecdote". The detector can't tell. It can only tell you which one reads more like a model.

Why AI detectors get it wrong

1. Careful, plain writing looks machine-made

Clear, well-edited prose is predictable on purpose. Lab reports, legal text, technical documents and anything written to a template score as AI more often, because the genre asks for predictable language.

2. Second-language writers get flagged more

In a 2023 Stanford study, seven detectors flagged on average 61.22% of TOEFL essays by non-native English speakers as AI, against about 5% of essays by US eighth-graders. A smaller vocabulary and safer sentence structures read as machine-like. More on who this hits in why does my writing get flagged as AI?.

3. Light editing hides AI text

The mirror image. A generated draft with a few sentences reworded, a personal detail added and the rhythm broken up often passes. Detectors catch careless AI use far more reliably than deliberate AI use.

4. Short and mixed texts

Below about 150 words there often isn't enough signal for a stable score, and an honest tool says so. A document that mixes a human introduction with an AI-drafted middle can confuse a single overall score. That's why sentence-level highlights matter more than the headline number.

How to read an AI detector result

A good detector gives you more than one number. Read them in this order:

If a tool gives you only a percentage, with no range, no confidence and no explanation, it's asking you to trust it blindly. Don't.

Using an AI detector fairly

If you're a writer or student, check your own work before you submit it, so nothing surprises you. If it flags, don't rewrite to beat the tool. That makes the writing worse. Keep your drafts and notes, because they're better evidence than any score.

If you're a teacher, editor or hiring manager, use the score to decide where to look, then talk to the person about their sources, drafts and choices. Five minutes of conversation tells you more than any percentage. Our guide to AI detector accuracy shows why even a 1% error rate can mean many wrong flags.

Frequently asked questions

How do AI detectors work?

They estimate how likely it is that a language model wrote a text, using word predictability (perplexity), sentence variation (burstiness), a classifier trained on labelled human and AI examples, or a hidden watermark. They read only the finished text, so the result is an estimate, not proof.

How do AI checkers work?

An AI checker is the same thing as an AI detector. It scores the text against patterns learned from human and AI writing and reports how AI-like it reads, often with highlighted sentences.

What is perplexity in AI detection?

Perplexity measures how surprising each word is given the words before it. AI text tends to have low perplexity because models pick likely words. Plain human writing can have low perplexity too, which is one source of false positives.

What is burstiness?

Burstiness is how much sentence length and structure vary through a text. Human writing usually varies more. AI writing tends to keep a steady rhythm.

How does GPTZero work?

GPTZero says it moved from perplexity and burstiness to a deep-learning model in autumn 2023. It now uses those two measures as one of seven indicators, alongside its deep-learning detection and a text search.

Do AI detectors use watermarks?

Some can. Google SynthID Text adds a watermark to Gemini output that a detector can check. Watermarks only work for models that add them, and they weaken when text is rewritten or translated.

Does Google penalize AI-generated content?

Google says it rewards helpful, high-quality content however it is produced. Using AI to mass-produce pages mainly to manipulate rankings breaks its spam policies. Thin, generic content is the problem, not the tool.

What is the difference between an AI detector and a plagiarism checker?

A plagiarism checker compares your text with existing sources and shows exact matches. An AI detector estimates the statistical style of your text. One shows where a sentence came from. The other guesses what kind of thing might have written it.