How do AI detectors work?

Updated October 31, 2026 · 7 min read

Short answer

Most AI detectors you can use today are trained classifiers: models trained on large sets of human and AI-written text until they learn the patterns in word choice and structure that tell the two apart. Turnitin, GPTZero and Pangram all describe their detectors this way. They read only the text in front of them and return a probability, not proof.

Two other approaches come mostly from research. Zero-shot methods such as DetectGPT and Binoculars measure how predictable a text is to a language model without training on AI examples. Watermarks are hidden statistical patterns an AI company adds when its own model writes: Google DeepMind added one to Gemini in 2024, while OpenAI said that year it had built one but was holding it back.

What are the main ways to detect AI text?

Method	How it works	Who uses it	Main weakness
Trained classifier	A model trained on labeled human and AI text learns to tell them apart	Turnitin, GPTZero, Pangram, Copyleaks, Sapling and most commercial tools	Weaker on models, topics and styles it wasn't trained on
Perplexity and burstiness	Measures how predictable each word is to a language model, and how much that varies	GPTZero's original 2023 model; the GLTR research tool (2019)	Plain, predictable human writing can look like AI
Zero-shot statistics	Compares how likely a text is under one or two language models, with no training on AI examples	Research methods: DetectGPT, Fast-DetectGPT, Binoculars	Paraphrasing defeats it; mostly tested on older models
Watermark	The AI company hides a statistical pattern in the words its model picks, then checks for it	Google DeepMind's SynthID-Text in Gemini; OpenAI built one but held it back	Works only on models that add it; rewording weakens it

How did AI text detection start?

The first idea was predictability. A language model writes by choosing likely next words, so AI text tends to be full of words a model would have predicted.

GLTR (2019) colored each word by how predictable a language model found it. In the authors' user study, this raised people's detection rate for generated text from 54% to 72%.
Perplexity and burstiness (GPTZero, 2023). GPTZero described perplexity as "a measure of how likely an AI model would have chosen the exact same set of words," and burstiness as "how much writing patterns and text perplexities vary over the entire document." Human writing tends to vary more.
Why it was replaced. Predictable writing isn't unique to AI. In a 2023 study, seven detectors of the time flagged essays by non-native English students far more often than essays by native speakers. When ChatGPT rewrote those essays with richer word choices, their perplexity rose and the average false positive rate fell from 61.3% to 11.6% (Liang et al.). GPTZero says it moved to a deep-learning model in autumn 2023, although an older passage on the same page still lists perplexity and burstiness as one of the current model's seven indicators.

How do most AI detectors work today?

Most commercial detectors are trained classifiers: the same kind of neural network that powers chatbots, trained instead to sort text into human and AI.

The basic recipe (2019). OpenAI fine-tuned a RoBERTa model as a classifier and caught text from its 1.5-billion-parameter GPT-2 about 95% of the time. Classifiers of that generation age badly: in a 2025 study, an open-source RoBERTa detector flagged 69% of human-written résumés, while GPTZero and Pangram flagged none at their best thresholds (Jabarian and Imas).
Turnitin says it uses "a modern deep learning approach based on the transformer architecture," trained on "both AI generated and authentic academic writing across geographies and subject areas." It scores each sentence and reports the share of sentences above a threshold.
GPTZero's 2026 technical report describes "a hierarchical, multi-task architecture enabling a flexible taxonomy of human and AI texts."
Pangram trains with what it calls "hard negative mining with synthetic mirrors": it pairs human texts with AI texts on the same topic and retrains on its mistakes, which it says cuts false positives. It says it reads "writing style, word choice, syntax choice, and grammatical structure."
Copyleaks lists signals such as frequency ratios, parts of speech and "syllable dispersion" alongside deep learning. Sapling scores the probability that each word or token is AI-generated and builds sentence scores from those.

None of these tools looks up sources or reads your editing history. Pangram says its detector "does not compare the text to other submissions in a database or on the web like traditional plagiarism checkers." That's a separate job: AI detectors vs plagiarism checkers.

What are zero-shot detectors?

Zero-shot methods need no training on AI examples. They use a language model's own probabilities as the test. Results below are from each paper's authors, on the models of the time. AUROC runs from 0.5 (a coin flip) to 1.0 (perfect).

DetectGPT (2023) found that "text sampled from an LLM tends to occupy negative curvature regions of the model's log probability function." It compares a passage with slightly altered copies of it. On fake news written by GPT-NeoX (20 billion parameters), it scored 0.95 AUROC, against 0.81 for the best earlier zero-shot method.
Fast-DetectGPT (2024) replaces the slow perturbation step with sampling. Its authors report about 75% better detection than DetectGPT, running about 340 times faster.
Binoculars (2024) contrasts two closely related language models. Its authors report it caught over 90% of ChatGPT samples at a false positive rate of 0.01%, without training on any ChatGPT data.
The catch: paraphrasing. In a 2023 test, an AI paraphraser cut DetectGPT's detection of GPT-2 XL text from 70.3% to 4.6% at a 1% false positive rate (Krishna et al.).

How do AI watermarks work?

A watermark is added when the text is generated, by the company that runs the model. Only that company's text carries it.

The "green list" method (2023). Before each word, the model picks a random set of "green" words and softly favors them. A checker that knows the secret can then spot the pattern statistically, from a short span, without access to the model (Kirchenbauer et al.).
Google DeepMind's SynthID-Text was tested on nearly 20 million live Gemini responses without a measurable drop in quality, and was added to Gemini (Nature, October 2024). Its authors note it needs the AI provider's cooperation, and that "generative watermarks are weakened by edits to the text, such as through LLM paraphrasing."
OpenAI said in 2024 that it had a text watermark that was "highly accurate," but that rewording with another model or running the text through translation made it "trivial to circumvention by bad actors," and that it "could stigmatize use of AI as a useful writing tool for non-native English speakers." As of that August 2024 update, it was holding it back.
Repeated paraphrasing dropped a watermark detector's detection rate from 99.3% to 9.7% at a 1% false positive rate in one study (Sadasivan et al.).
The law. The EU AI Act requires providers of AI systems that generate text to mark outputs "in a machine-readable format" so they are "detectable as artificially generated." Systems on the market before 2 August 2026 must comply by 2 December 2026. There's an exemption for "an assistive function for standard editing," and the Act doesn't require anyone to run detectors on people's writing.

So a watermark can show that a text came from a model that marks its output. It can't show that a text without one is human-written.

Why do AI detectors make mistakes?

There's a ceiling. Sadasivan et al. tie the accuracy of the best possible detector to how different human and AI text really are. As models write more like people, detection gets harder.
Short texts. There's less pattern to measure. QuillBot needs at least 80 words, Originality.ai 100, Turnitin 300, and Copyleaks recommends at least 350. OpenAI's own 2023 classifier was "very unreliable on short texts (below 1,000 characters)." It caught only 26% of AI text in OpenAI's own test and was withdrawn in July 2023 for low accuracy.
New models and settings. The RAID benchmark (2024) found "a clear bias towards domains and models that the detectors have trained on." How detectors do on Claude, Gemini and other models.
Rewriting. Paraphrasers and humanizers are built to change the patterns detectors look for. What the research shows.
Light AI editing. In a 2026 Notre Dame study, human abstracts lightly edited by AI were flagged 38 to 80% of the time, depending on the tool and era.
Writing style. Plain or formulaic writing, including some non-native English writing, can look predictable. The research on non-native writers.

What does this mean when you read a result?

A detection score is a model's estimate about patterns in the text, not a record of how it was written. What scores mean in each tool.
Longer, unedited passages give more reliable results than short or heavily edited ones.
Look at which passages were flagged, not just the overall number. AI Checker Plus uses a trained detection model and labels each passage AI-generated, AI-assisted or human-written, so you can see what drove the result.
Treat a result as one piece of evidence, next to drafts, version history and a conversation with the writer.

Questions people also ask

Do AI detectors use perplexity and burstiness?

Some did. GPTZero's original 2023 model measured perplexity (how predictable the words are to an AI model) and burstiness (how much that varies across a document). GPTZero says it moved to a deep-learning model in autumn 2023. Most current detectors, including Turnitin's, are classifiers trained on human and AI text.

Can an AI detector see how I wrote my text?

No. A detector reads only the text you give it. Pangram, for example, says its detector does not compare text with a database or the web and does not capture writing-process data such as cut-and-paste. Drafts and version history are separate evidence that a detector can't see.

Does ChatGPT put a watermark in its text?

The last public statement we found from OpenAI, an August 2024 update, said it had built a text watermark but was holding it back, partly because translation or rewording with another model defeats it. Under the EU AI Act, AI systems already on the market before 2 August 2026 must mark AI-generated text in a machine-readable way by 2 December 2026, so this may change.

Why do AI detectors flag human writing?

Because they judge statistical patterns, not how a text was actually written. Short passages, plain and predictable writing, some non-native English writing and human text lightly edited with AI tools can all look like AI output to a model. That's why every major vendor says a score shouldn't be the only evidence.

Sources

By the AI Checker Plus team. We research and draft these guides with help from AI tools, then check every claim against the sources listed above.