AI Detectors Explained: How GPTZero, Turnitin & Originality Work

AI detectors do not actually know whether a human or a machine wrote your essay. They make a statistical guess, and they are wrong more often than their marketing admits. Understanding how that guess is made tells you everything about when to trust it, which is rarely, and why so many real students get flagged for work they wrote themselves.

Here is what is happening under the hood of GPTZero, Turnitin, Originality.ai, and every other detector, in plain English.

What an AI detector actually does

An AI detector is a pattern matcher. It does not read for meaning or check facts. It measures statistical properties of your text and compares them against what AI writing tends to look like, then outputs a probability that the text was machine-generated.

That output is an estimate, not proof. Every major detector says so in its own documentation. The number on the screen is the tool's confidence, not a fact about who wrote the words.

Detectors lean on three things: perplexity, burstiness, and trained classifier models.

Signal 1: Perplexity

Perplexity measures how predictable the text is. Language models write by choosing the most probable next word, over and over, so their output tends to be smooth and unsurprising. That reads as low perplexity.

Human writing is messier. People make odd word choices, take detours, and phrase things in ways a model would consider unlikely. That unpredictability reads as higher perplexity. So when a passage is unusually smooth and predictable, a detector treats it as a sign of AI, according to GPTZero's own explanation of how AI detectors work.

Signal 2: Burstiness

Burstiness measures how much sentence length and structure vary across a passage. It is often calculated as the variation in sentence length across a document.

Humans write in bursts. A long, winding sentence with several clauses, then a short one, then a three-word fragment. AI tends to produce sentences of similar length and shape, paragraph after paragraph. That uniformity is low burstiness, and detectors read it as machine-like.

This is also why varying your sentence length is the single most effective way to make writing read as human. It is the exact metric the tool is watching.

Signal 3: Trained classifiers

The better detectors no longer rely on perplexity and burstiness alone. They run your text through a classifier model, usually a fine-tuned transformer such as RoBERTa or DeBERTa, that has been trained on large sets of human and AI writing. The classifier weighs many features at once, including the two signals above plus stylometric patterns, and produces a single probability score.

Modern tools combine all of this rather than betting on one metric. GPTZero, for example, has expanded to a multi-component system. But more signals do not mean more certainty, and that is the catch.

What about watermarking?

A different approach is to have AI models embed an invisible statistical watermark in their output that a detector can later read. In theory it is more reliable than guessing. In practice it barely helps, because it only works if every model watermarks its output the same way. 2026 research on the topic is blunt: without shared standards and universal adoption, watermarking does not provide reliable, externally verifiable detection, and it is easy to defeat with light editing. No detector you encounter today is catching AI writing through watermarks.

Why detectors get it wrong so often

Here is the core problem. The patterns detectors associate with AI, smooth predictable wording and uniform sentence structure, also appear naturally in plenty of human writing.

• Non-native English writers get hit hardest. A 2026 analysis found detectors flagged 61.3% of TOEFL essays by non-native writers as AI, versus 5.1% for native writers, because second-language writing tends toward simpler vocabulary and uniform structure. The bias is well documented.
• Technical and formal writers trip detectors because clear, structured prose has naturally low perplexity.
• Edited or paraphrased text slips through in the other direction, because changing the patterns lowers the score.

A 2026 study in the International Journal for Educational Integrity evaluating these tools in academic settings found they are not accurate or reliable enough to serve as evidence. Broader 2026 research reaches the same place: detection is an arms race with no tool achieving dependable accuracy, especially on edited or non-native text. As one 2026 paper puts it, the field is moving beyond detection toward rethinking assessment entirely, precisely because the tools cannot be trusted on their own.

How the major detectors compare

They differ mostly in how aggressively they flag.

• GPTZero pioneered the perplexity and burstiness approach and now uses a multi-component model. It is widely accessible, which means anyone can run your work through it.
• Turnitin is conservative by design and hides scores below 20% as unreliable, but it is heavily biased against non-native English writing. (More on whether Turnitin detects ChatGPT.)
• Originality.ai is the most aggressive, built for publishers who would rather over-flag than miss anything, which produces more false positives.

Different tools give different verdicts on the same text. That alone tells you how much weight a single score deserves.

What this means for you

Treat any AI detector score as a weak signal, never a verdict. If you are being judged by one, know that it is a probability estimate from a tool that independent research repeatedly finds unreliable. And if your natural writing happens to trip detectors, whether because of your language background or a clean, formal style, rewriting for natural variation and voice is a legitimate way to read as the human you are, and our guide covers how to humanize AI text step by step. That is the same work a humanizer like BlueHumanizer does: it changes the exact signals these tools measure, so genuine human-sounding writing is read as human.

Can you beat an AI detector?

Yes, and how you do it reveals exactly how fragile these tools are. Because detectors score statistical patterns rather than read for meaning, anything that changes those patterns changes the verdict. Vary your sentence length, raise the unpredictability of your phrasing, rebuild the structure, and the score drops. Not because anything was tricked, but because the text no longer matches the pattern the tool was trained on.

That is the uncomfortable truth sitting under every detector. The same edit that helps a real student avoid a false accusation also helps anyone evade detection. A tool that can be defeated by ordinary editing was never solid enough to accuse someone on its own. Researchers keep arriving at the same conclusion: detection is a moving target, and every advance on the detection side is quickly undone on the writing side.

For most people, the honest use is defensive. If your natural writing reads as predictable, whether because English is your second language or because you write in a clean, formal style, you are at genuine risk of a false flag from a tool that gets it wrong often. Reshaping your text so it carries the natural variation of real human writing is a reasonable response to a flawed system. That is what a humanizer does: it adjusts the exact signals detectors measure so that honest, human-written work is read as honest and human.

None of this makes a detector score meaningless. It makes it a starting point for a conversation, never the end of one.

The bottom line

AI detectors work by measuring predictability and sentence variation and running the result through a trained classifier. They produce a probability, not proof, and they are wrong often enough that major universities have stopped trusting them. Knowing how the guess is made is the best defense against being judged by it.