How do AI detectors work?

How do AI detectors work?

Every AI text detector starts from the same fact: a large language model writes by choosing a likely next word, over and over. Turnitin's own FAQ puts it plainly. Models such as ChatGPT "tend to generate the next word in a sequence of words in a consistent and highly probable fashion," while human writing "tends to be inconsistent and idiosyncratic." Detectors try to measure that difference. There are three main ways to do it.

Perplexity and burstiness

Perplexity is a measure of how surprised a language model is by each word. If the model could easily have guessed the next word, perplexity is low. If the word is unexpected, perplexity is high. Liang and colleagues describe it this way in their 2023 paper and note that most GPT detectors use text perplexity to decide whether text is AI-written.

Burstiness is how much perplexity changes across a document. GPTZero's founder explained in an early technical note that burstiness is "a measure of how much the perplexity varies over the entire document," because people mix short and long, plain and unusual sentences while models tend to keep a steady rhythm. That note now carries an update: as of autumn 2023, GPTZero no longer uses perplexity and burstiness and moved to a deep learning architecture; my GPTZero review covers the current product and its plans. Many free checkers still rely on simple statistics like these, and the ZeroGPT vs GPTZero comparison shows how far apart two similarly named detectors can land on the same human essays.

Example. Here is a sentence I wrote to sound like a default chatbot answer, followed by a version I rewrote by hand. Neither came from a tool.

• Stiff: "Effective time management is essential for achieving success in both academic and professional settings."
• Natural: "I lost two weekends in March to a deadline I saw coming in January. That is when I started blocking out Sunday mornings."

The first sentence is built from the most expected words in the most expected order, so a language model finds it very easy to predict: low perplexity. The second has a specific month, a personal detail and an uneven rhythm, so each next word is harder to guess. That gap is what perplexity-based detectors measure. It is also why a careful human who writes the first kind of sentence, as many students are taught to, can be flagged.

My take: perplexity is a reasonable signal of style, not of authorship. Plain, correct, template writing scores low whoever wrote it.

A research version of the idea is DetectGPT. Mitchell and colleagues at Stanford found that text sampled from a model tends to sit in a particular region of that model's probability function, and they built a test that needs no training data. Their 2023 paper reported an AUROC of 0.95 for detecting fake news articles from a 20 billion parameter model, up from 0.81 for the best earlier zero-shot method. AUROC is a measure of how well a test separates two groups, where 1.0 is perfect.

Trained classifiers

Most commercial detectors today are classifiers: a model trained on large sets of text labeled human or AI, which learns its own patterns rather than following a fixed rule. Turnitin says its current system is a transformer-based deep learning model trained on its in-house corpus of academic writing. It also says the model "is not explicitly programmed to evaluate specific signals such as burstiness, perplexity, or other individual metrics," and that individual predictions "may not always be explainable in simple feature-by-feature terms."

In practice, Turnitin's AI writing detection FAQ describes the steps like this, as read on 29 October 2026:

1. The model drops text that is not prose: lists, headers, references, tables, formulas and code.
2. It gives each remaining sentence a score between 0 and 1, in the context of the sentences around it.
3. Sentences above a set threshold are labeled AI and the rest are labeled human.
4. The document score is the percentage of sentences above the threshold. Turnitin does not display scores below 20%.

The FAQ lists 33 language models the English detector covers, from Claude, Gemini, GPT, DeepSeek, Grok, Llama, Mistral and Amazon Nova families, released between December 2023 and March 2026. Turnitin says it updated the English model architecture in October 2026. My Turnitin AI detection review covers the report, its thresholds and the appeal process in detail.

A classifier is only as good as the text it learned from. OpenAI's own 2023 classifier warned that neural network classifiers "are known to be poorly calibrated outside of their training data" and can be very confident and wrong on unfamiliar input.

My take: a classifier is the strongest of the three methods on the kind of text it was trained on, and the hardest to argue with when it is wrong, because no one can point to the feature that tipped it.

Watermarks

A watermark is different: the AI company changes how its own model picks words so that the output carries a hidden statistical pattern, then checks for that pattern later. Kirchenbauer and colleagues at the University of Maryland proposed a well-known scheme in 2023 that softly promotes a randomized set of "green" tokens during generation and tests for them with a statistical test.

Google DeepMind published SynthID Text in Nature in October 2024. The paper reports a live experiment on nearly 20 million Gemini responses that found no loss of text quality. Google's developer documentation says SynthID Text is open source, that detection is probabilistic with three possible results (watermarked, not watermarked or uncertain), and that it is robust to some changes such as cropping or mild paraphrasing but "not designed to directly stop motivated adversaries."

OpenAI said in an August 2024 update that it had built a text watermarking method that was "highly accurate" against local edits but less robust against rewording with another model or translation. It also said the method could stigmatize AI use as a writing tool for non-native English speakers, and it has not shipped a text watermark.

The key limit: a watermark detector can only recognize text from the model that added the watermark. It says nothing about text from any other tool.

The three methods side by side

Here is how I would compare the three main approaches if I were choosing which result to trust most.

My take: if a school or employer shows you a score, the first question I would ask is which of these three methods produced it. The answer tells you most of what that score can and cannot mean.

Are AI detectors accurate?

Not reliably enough to prove authorship on their own. Accuracy depends on the detector, the text, its length, the writer and whether anyone edited it. Vendor figures are measured on vendor test sets; independent studies measure on their own samples and often find much lower numbers.

What vendors say. Turnitin says it aims to keep the false positive rate, meaning human text wrongly flagged as AI, under 1% for documents with over 20% AI writing. To hold that rate, it accepts misses: "if we identify that 50% of a document is likely written by an AI tool, it could contain as much as 65% AI writing." Before each model update, Turnitin says it tests on over 700,000 academic papers written before ChatGPT.

What OpenAI found with its own tool. OpenAI's January 2023 classifier correctly flagged 26% of AI-written text and wrongly labeled human text as AI 9% of the time. OpenAI withdrew it on 20 July 2023 "due to its low rate of accuracy."

What independent studies found:

The Weber-Wulff team's study of 14 detection tools also found a "main bias towards classifying the output as human-written," so the tools more often missed AI text than wrongly accused people. Perkins and colleagues found the other risk in their human control group, where only 67% of tests were accurate. Both results point to the same conclusion: error rates are large enough to matter for real decisions.

My take: I read vendor accuracy figures as a best case on the vendor's own test set, and independent study figures as a warning about what happens outside it. The truth for your text sits somewhere in between, and no one can tell you exactly where.

How accurate are AI detectors on different kinds of writing?

Accuracy is not one number. The same detector can do well on one kind of text and badly on another.

• Non-native English writing. In Liang and colleagues' study, at least one of seven detectors flagged 97.8% of the human-written TOEFL essays, and all seven flagged 19.8%. When the authors used ChatGPT to enrich the essays' vocabulary, false positives fell sharply, which shows the detectors were reacting to word choice.
• Short text. Turnitin does not process documents with under 300 words of qualifying prose, and it says that in documents of only a few hundred words the prediction is "mostly all or nothing." OpenAI said its classifier was "very unreliable on short texts (below 1,000 characters)."
• Formulaic writing. Turnitin's FAQ says false positives can include "content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas."
• Lists, tables and code. Turnitin skips them, and OpenAI said its classifier was unreliable on code.
• Edited or paraphrased AI text. Detection rates drop, sometimes sharply. Detectors respond by training on edited text, as covered below.

My guide to AI detector false positives walks through these risks with a worked example and five checks to run before a score drives any decision.

Is there a 100% AI detector?

No. No published study has found one, and the vendors do not claim one. Turnitin says its percentage "should not be used as the sole basis for action," OpenAI wrote that "it is impossible to reliably detect all AI-written text," and the 14-tool Weber-Wulff study found none above 80% accuracy.

There is also a structural reason. A detector has to pick a threshold. Set it strictly and it misses more AI text; set it loosely and it flags more human text. Turnitin openly chose the strict side. Sadasivan and colleagues at the University of Maryland went further in their Transactions on Machine Learning Research paper, Can AI-Generated Text be Reliably Detected?, showing that repeated paraphrasing reduced detection rates across watermark, classifier, zero-shot and retrieval detectors with only slight loss of text quality. They also showed that watermarks can be spoofed so that human text is flagged as AI.

A vendor's marketing page that promises perfect accuracy is making a claim no independent test has supported. If a detector advertises 99% or 100% accuracy, I'd treat that as a reason to trust it less, not more. My AI detector rankings compare documented limits rather than headline accuracy numbers.

Can you make AI text undetectable?

No method makes text reliably undetectable, and the question usually hides a different one. Here is what the research and the vendors show, without the how-to.

Detection drops after editing, but not to zero and not forever. Krishna and colleagues showed in their NeurIPS 2023 paper, Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense, that heavy paraphrasing fooled several detectors of that time. The same paper showed that a retrieval defense, where the AI company checks text against its own log of outputs, caught 80% to 97% of the paraphrased text while flagging 1% of human text. Detection is an arms race in which both sides keep changing.

Detectors now train on edited text. Turnitin added AI paraphrasing detection in December 2023 and, on 27 August 2025, announced AI bypasser detection designed to flag text modified by humanizer tools, in English. Turnitin says it will not name the tools it detects. My guide on whether Turnitin can detect humanized AI text covers that update in detail. So a tool that slipped past a detector in one month can be caught the next, and old submissions can be rescored if they are resubmitted.

Edited text can read worse. Perkins and colleagues found that some manipulation techniques made AI text almost undetectable but produced writing that "would very likely receive poor marks in a real-world setting because of the high number of errors."

Detectors are not the only check. Teachers compare work with earlier writing, ask students to explain their argument, and look at version history. A detector score is only one input. My guide on whether teachers can detect AI covers those checks.

If your goal is better writing, the useful questions are about quality and permission. Editing an AI-assisted draft so it reads naturally, in your own voice, is ordinary editing where AI use is allowed. Using any tool to hide AI use that a course or employer has banned is an integrity problem, whatever the detector says. My guide to humanizing versus paraphrasing explains what those tools actually change in a text.

My take: the people who get the most from AI writing tools are the ones using them where they are allowed, for marketing copy, emails, SEO drafts or smoothing a second language, and editing the result properly. For that work the detector question barely matters; the reader is the test.

Does bypass AI work?

"Bypass" tools are AI humanizers marketed on the claim that their output passes detectors. That claim belongs to the vendor. Independent research gives a mixed and moving picture:

• Some edits lower detection rates in studies. Perkins and colleagues found accuracy fell by a mean of 17.4% when simple manipulation techniques were applied, from 39.5% to 22.1%.
• Detectors adapt. Turnitin's August 2025 bypasser update exists because of those tools, its release notes list another model update on 12 February 2026, and its FAQ describes a new English model architecture in October 2026. Results measured against an older version say little about the current one.
• Universities warn about the tools themselves. The University of Melbourne tells students that many free detector sites are "aimed at exploiting student anxieties" by flagging human work so they can sell humanizer tools, which "usually use AI and are often picked up by the University's detector tool." It also warns that pasting work into free tools can give away your intellectual property.

So the honest answer is: sometimes, against some detectors, for a while, with no guarantee. Schools set their own AI rules, and passing a detector does not make a banned use allowed. My guide to do AI humanizers work reviews the studies on both readability and detection, and my ranking of AI humanizers compares the tools on price, languages and privacy terms rather than on detector claims.

What a detector score does and doesn't prove

A detector score is a model's estimate. It is useful to read it for what it is.

At best it tells you where a model found patterns it associates with AI text. What it cannot tell you:

• Who wrote the text, or how much help they had.
• Whether AI use was allowed for this task.
• Whether a low score means the text is human. Turnitin says a 50% score could hide 65% AI writing, and OpenAI's classifier caught only 26%.
• Anything reliable about short text, lists, code or languages the detector does not support.

Turnitin's own guidance says it "does not make a determination of misconduct" and that the percentage should not be the sole basis for action. Vanderbilt University went further in August 2023 and disabled Turnitin's AI detector, saying: "we do not believe that AI detection software is an effective tool that should be used." Curtin University in Australia disabled Turnitin's AI detection from 1 January 2026 while keeping text-matching checks on.

The fairest evidence is about process: drafts, version history, notes, sources and the writer's ability to explain the work. If your question is about Turnitin specifically, my guide to whether Turnitin checks for AI covers what its check includes, my Turnitin similarity score guide explains how the separate text-matching score works, and the GPTZero vs Turnitin comparison sets out how the two most common detectors differ. Every guide in this series sits in the HumanRank guides hub.

A quick checklist before you trust a score

My take: fail any of the first three rows and I would ignore the score.

How we checked this

• Read on 29 October 2026: Turnitin's AI writing detection FAQ, its AI writing model release notes and its 27 August 2025 bypasser press release; GPTZero's note on perplexity and burstiness; OpenAI's January 2023 classifier post with its July 2023 update and its May 2024 provenance post with the August 2024 update; Google's SynthID Text developer page.
• Studies read: Liang et al. (Patterns, 2023), Weber-Wulff et al. (International Journal for Educational Integrity, 2023), Perkins et al. (International Journal of Educational Technology in Higher Education, 2024), Krishna et al. (NeurIPS 2023), Sadasivan et al. (TMLR), Mitchell et al. (ICML 2023), Kirchenbauer et al. (2023) and Dathathri et al. (Nature, 2024).
• University guidance: Vanderbilt University (August 2023), Curtin University (September 2025 notice) and the University of Melbourne's student advice.
• Method: I attributed every figure to its source and year, and kept my own judgement in clearly marked "My take" lines. The site's general approach is on the HumanRank ranking methodology page.

Frequently asked questions

How do AI detectors know if something is AI-written? They do not know; they estimate. Most compare the text with patterns learned from known AI and human writing, and give a probability. A few check for a watermark the AI company added.

What is perplexity in AI detection? Perplexity measures how easily a language model can predict each next word. Low perplexity means predictable text, which many detectors read as a sign of AI, even though careful human writing can be predictable too.

Can AI detectors be wrong about human writing? Yes. Turnitin aims for under 1% on documents, OpenAI's 2023 classifier mislabeled 9% of human text, and one 2023 study found a 61.3% average false positive rate on non-native English essays.

Should schools rely on AI detectors? That is a policy choice. Turnitin says its score should not be the sole basis for action, and Vanderbilt and Curtin chose to disable Turnitin's AI detection altogether.

REVIEW NOTEI checked this guide on 29 October 2026 against Turnitin, GPTZero, OpenAI and Google documentation, and against seven peer-reviewed or preprint studies on AI text detection. Every number on this page belongs to the named source, and every opinion is mine.