How Do AI Detectors Work?
What detectors measure, why they sometimes flag human writing, and how to read the number they give you.
By the AI Checker Humanizer team. Published . Updated .
The short answer
An AI detector reads a piece of text and estimates how likely it is that a language model wrote it. The methods range from older statistical measures, perplexity and burstiness, to trained classifiers, which is how GPTZero and Turnitin describe their tools today.
- Predictability. Language models tend to choose likely words, so AI detectors measure how predictable a text is to a particular model. This is usually called perplexity.
- Variation. How much sentence length and structure change through a text. This is usually called burstiness. An even rhythm turns up in both human and AI writing; detectors only treat it as one signal among others.
- Trained classifiers. A model is shown large amounts of human and AI writing and learns to tell them apart. GPTZero and Turnitin both describe their tools this way.
Watermarks are a separate thing. A watermark is a hidden statistical signal that some AI companies build into their own output, and only tools with the right key can check for it. Ordinary detectors don't read watermarks at all.
Perplexity and burstiness, explained with examples
Here are two short paragraphs about the same thing. We wrote both for this article to show the difference; neither is from a real student.
Remote work has transformed the way organizations operate. It offers employees greater flexibility while reducing overhead costs for employers. However, it also presents challenges related to communication and collaboration. Ultimately, the success of remote work depends on clear policies and strong leadership.
We went remote in March and never really came back. I don't miss the commute. I do miss being able to lean over and ask Priya a question without booking a call, which sounds small until you're three Slack threads deep on a Friday.
Perplexity measures how predictable a text is to a particular language model: the lower it is, the less the model is surprised by each word. The first paragraph is built from stock phrases such as “greater flexibility” and “challenges related to”, the kind of wording models produce all the time. The second has specific details like “three Slack threads deep”. To compare their actual perplexity you'd score both with the same model; these two are only meant to show the idea.
Burstiness is about variation. The first paragraph has four sentences of roughly the same length and shape, each one making a balanced point. The second has sentences of 10, 5 and 29 words, a much more uneven rhythm. Either a person or a model could write either paragraph; AI detectors only see that one pattern is more common in AI output than the other.
Neither measure is reliable on its own. A careful human writer producing a formal report can have low perplexity and low burstiness. That's one reason detectors that relied only on these two numbers flagged so much human writing.
Classifier models: how most detectors work today
Most current AI detectors are classifiers. The builders collect a large set of texts with known authors, some human and some from AI models, and train a model to separate the two. The model learns patterns that are hard to list by hand: how ideas are ordered, which hedges get used, how paragraphs open and close.
Two design choices make a big difference to what you see.
- Whole document or sentence by sentence. Some tools give one score for the whole text. Others score each sentence or segment and then combine them, which lets them highlight the parts that look most like AI. Sentence scores are noisier, because a single sentence carries less evidence.
- Where the threshold sits. Every detector has to pick the score at which it calls something AI. A low threshold catches more AI text and wrongly flags more human writing. A high one does the opposite. Two tools can use similar models and still disagree simply because they set the line in different places.
Training data matters too. A classifier learns from text produced by the models that existed when it was trained. When a new model comes out with different habits, a detector can miss it until the builders add samples from that model and retrain, so it's worth checking which models a detector has been tested against, and when.
GPTZero, one of the best-known tools, used perplexity and burstiness when it launched in January 2023. It says it switched to a deep learning model that classifies text sentence by sentence later that year. Turnitin describes something similar: a model scores each sentence in the context of the text around it, and the document score is the share of sentences above its threshold. We only know what these companies publish about their own tools.
Watermarks are a different thing
A watermark works at the moment the text is written. When a model chooses between several equally good words, a secret key nudges which one it picks. Nothing visible is added and there are no hidden characters, but someone holding the key can later run a statistical test and see the pattern.
- Anthropic says recent Claude models, including Fable 5.1 and the 5.5 versions of Opus, Sonnet and Haiku, add a text watermark. Older ones such as Sonnet 4.5 don't. Its detection is in private preview for groups like educators, researchers and regulators. Our Claude AI detector page has more on this.
- Google watermarks text from the Gemini app and website with SynthID and has open-sourced the method. Its public SynthID Detector checks images, video and audio; we couldn't confirm that it checks text for the public.
- OpenAI says verification of its text marks is available only to approved organizations, such as research and academic institutions.
Ordinary AI detectors, ours included, can't see these watermarks because they don't have the keys. They judge the writing itself. As Anthropic puts it, a detected mark is a signal that Claude processed the text, not final proof, and a missing mark doesn't mean a person wrote it. (Sources checked November 11, 2026.)
Copied AI text can also contain zero-width spaces and other invisible Unicode characters. Unlike a statistical watermark, these can be found and inspected with a tool, and their presence alone doesn't prove that a text was AI-written. Our hidden character cleaner shows and removes them.
Which words trigger AI detection?
There's no list of banned words. A detector looks at the whole piece of writing, so swapping out a few words rarely changes the result. That said, some phrases do turn up far more often in AI text than in human writing.
We counted phrases in our own test collection: 150 texts written by a recent GPT model (about 103,000 words) and 1,299 human texts of the same kinds, academic writing, blog posts and emails (about 1.09 million words).
| Phrase | GPT, per 100,000 words | Human, per 100,000 words |
|---|---|---|
| I would appreciate | 23.1 | 1.1 |
| not simply | 20.7 | 0.3 |
| should therefore | 17.7 | 0.6 |
| best understood as | 10.3 | 0 |
| rather than merely | 10.3 | 0 |
| does not automatically | 9.3 | 0 |
| furthermore | 2.8 | 15.3 |
| crucial | 4.7 | 6.4 |
| delve | 0 | 0.2 |
In our GPT samples, the more frequent phrases included contrasts such as “not simply” and “rather than merely”, conclusions using “therefore”, and polite email requests. The bottom is the old list people still quote. “Delve” didn't appear once in the GPT texts, and “furthermore” was more common in the human writing. Our ChatGPT humanizer page has a longer list, and how to humanize AI content covers editing by hand.
The same goes for em dashes. In our collection the GPT texts used about 0.5 per 1,000 words, against about 1.0 in the human writing. Habits shift from one model version to the next, which is why AI detectors that rely on word lists age quickly.
Why detectors flag human writing
A false positive is human writing that a detector calls AI. They happen more often with some kinds of text.
- Short texts. A paragraph or two doesn't give a detector much to go on. Ours won't give a verdict on anything under 150 words (200 in Spanish, Portuguese and French).
- Formulaic writing. Lab reports, cover letters and five-paragraph essays follow fixed patterns, and fixed patterns are what AI produces too.
- Writing in a second language. Careful writers working in a second language often choose common words and safe sentence shapes. A 2023 Stanford study tested seven AI detectors on 91 TOEFL essays by non-native English speakers and found they wrongly flagged 61% of them on average.
- Heavily edited text. Running your own draft through grammar or rewriting tools can push it toward the smooth, even style detectors associate with AI.
The rate depends a lot on the detector. In our English tests, our detector flagged 1 of 1,096 human texts and none of 600 essays by non-native English writers. Other tools publish different numbers, and some publish none.
How detectors handle essays
Student essays are one of the hardest cases. Many assignments ask for a set structure, a neutral tone and a summary of well-known material, which is exactly what AI writes well. Students also polish their drafts, sometimes with tools.
A detector score alone doesn't establish how an essay was written. Drafts, sources and an explanation of the writing process give a teacher much more to go on, and many schools' own rules say a score can't be the only evidence. If you want to see how your own essay reads before you submit it, the AI essay checker shows which sentences look like AI.
How our detector works
Our detection engine scores every sentence with an AI model, reading each one alongside the text around it. It combines those scores with 12 writing measurements, such as stock phrases, how often transitions appear, sentence and word length, and how varied the vocabulary is.
We set the thresholds on our own collection of human and AI writing, then checked the result on texts the detector had never seen. These are the English results:
| Texts | How many | Result | 95% bound |
|---|---|---|---|
| Human-written texts | 1,096 | 1 flagged as likely AI (0.09%) | At most 0.52% |
| Essays by non-native English writers | 600 | 0 flagged as likely AI | At most 0.64% |
| AI-generated texts from 6 model families | 278 | 270 caught as likely AI (97.1%) | At least 94.4% |
Each sentence gets a color, so you can see where the score comes from: red sentences read like AI, yellow ones are unclear, and grey text wasn't scored. The result says whether text reads like AI, not which model wrote it; telling GPT, Claude and Gemini apart reliably is a harder problem that we don't claim to solve.
You can try it below, or on the full AI detector.
0 / 2,000 words
Frequently asked questions
Can AI detectors be wrong?
- Yes, in both directions. They can flag human writing, especially short, formulaic or very careful text, and they can miss AI text that someone has rewritten. A score is an estimate, not proof.
What is the 30% rule in AI writing?
- There isn't one. Some schools and publishers set their own limit on how much AI help is allowed, and some use a detector percentage as a trigger for a closer look. Check the rules where you're submitting.
Can professors tell if I used ChatGPT?
- Sometimes. Many use a detector, and many also notice when an essay doesn't sound like the student's other work, cites sources that don't exist, or can't be explained in conversation. Following your course's rules on AI use is the safer path.
Do AI detectors work in other languages?
- Some do, but most are tested mainly on English, and accuracy in other languages is usually lower or unpublished. Ours scores English, Spanish, Portuguese and French and adds a language only after it passes our tests.
Can AI detectors catch paraphrased text?
- Sometimes. Rewriting can make AI text harder to detect, but results vary a lot by detector and by how the text was rewritten.
How does Turnitin's AI detection work?
- Turnitin scores each sentence in context with its own model and reports the share of sentences it labels as AI. It's only available if your school licenses it and switches it on. More in our article on Turnitin AI detection.
Are AI detectors accurate?
- Accuracy varies by tool and by text. Long, unedited AI text is generally easier to detect than short or rewritten passages, and false positive rates on non-native writing differ a lot between detectors. See what the research says.