What AI detector does turnitin use?
Turnitin uses its own in-house AI writing detector, not GPTZero, Copyleaks, or any third-party engine. Launched in April 2023, it is a transformer deep-learning model trained on academic writing, scores every sentence from 0 to 1, requires at least 300 words of English prose, and reports one percentage inside the Similarity Report.
Why — the first-principles explanation
Turnitin built its detector in-house because it owns something no startup has: a giant archive of student papers written before ChatGPT existed. A detector is only as good as the "human writing" it was trained to recognize, and Turnitin could train on decades of real essays across subjects, countries, and skill levels — including second-language learners, whom the company says it deliberately over-sampled to reduce bias.
The machinery is a transformer — the same architecture that powers the language models it hunts. That is not a coincidence. Language models generate text by repeatedly picking a likely next word, which leaves a statistical fingerprint: the word sequence is unusually smooth and predictable compared with human prose. Older detectors measured this with two hand-built numbers, perplexity (how surprising the words are) and burstiness (how much sentence length varies). Turnitin's whitepaper says a transformer beats those hand-built measures because it can catch subtler, longer-range patterns that perplexity alone misses.
Mechanically, the model slides a window of a few hundred words across your document, scoring each window between 0 (almost certainly human) and 1 (strongly plausible AI). Each sentence gets a weighted average of the windows it appears in, and is flagged if that average clears a threshold — typically between 0.8 and 1. The document-level number you see is simply the share of sentences that got flagged. That is why Turnitin needs 300+ words: fewer sentences means an unstable average.
One detail matters more than the brand name: Turnitin refuses to publish an "accuracy" figure at all. Its whitepaper says accuracy is too easy to game — a detector that labeled everything human would score 99% accurate on a realistic dataset. Instead it publishes recall (how much AI it catches) and false positive rate (how often it accuses an innocent paper), which are the two numbers that actually decide whether a student gets called into an office.
An example that makes it click
Think of a piano teacher who has heard 800,000 student recitals over thirty years. She has never seen a player piano, but the first time one rolls in, she notices instantly: every note is perfectly on the beat, every phrase is the same length, nothing rushes or drags. She cannot point to one wrong note — the whole thing is too even.
That evenness is what Turnitin's model listens for. And notice the catch built into the analogy: a very disciplined, very careful human student who plays metronomically can sound like the player piano. The teacher is not reading minds; she is judging texture. That is exactly why Turnitin tells instructors the score is a conversation starter, not a verdict.
Key facts
- Turnitin's AI writing detector is proprietary and built in-house; it launched in April 2023 and is not licensed from GPTZero, Copyleaks, ZeroGPT, or Originality.ai.
- The system is a transformer deep-learning model that slides a segment window of roughly a few hundred words across a document, scoring each window from 0 to 1 (Turnitin whitepaper, August 2023).
- Minimum document length is 300 words; the model is tuned for English only, per Turnitin's published architecture whitepaper.
- A document is labeled AI-written only if more than 20% of its sentence-level scores clear the model threshold — Turnitin found a higher incidence of false positives below 20%.
- On a stress test of 800,000 papers submitted before 2019 (pre-GPT-3), Turnitin reported a document-level false positive rate of 0.7% and a sentence-level rate of 0.2%.
- Turnitin explicitly declines to report an 'accuracy' metric, arguing it is too easily manipulated; it reports recall and false positive rate instead.
▶ The 60-second explainer (script)
Turnitin doesn't buy its AI detector from anyone. It built its own, in-house, and launched it in April 2023. Here's what's actually under the hood. It's a transformer — the same type of model that powers ChatGPT — but trained backwards: trained to recognize genuine student writing from decades of real submissions, including work by second-language learners. It works by sliding a window of a few hundred words across your paper. Each window gets a score from zero to one. Zero means almost certainly human. One means strongly plausible AI. Every sentence gets the average of the windows it sits in, and the percentage you see on screen is just the share of sentences that got flagged. That's why Turnitin needs at least three hundred words, and why it only works in English. Now the number that matters. Turnitin ran the detector across eight hundred thousand papers submitted before 2019 — before GPT-3 existed, so all human. It flagged 0.7 percent of them as AI-written. That's the false positive rate, and it's not zero. Which is why Turnitin itself says the score is a starting point for a conversation, not proof of anything.
What authoritative sources say
People also ask
Does Turnitin use GPTZero or Copyleaks?
No. Turnitin's AI writing detection is entirely its own model, developed internally and trained on its own archive of academic writing. GPTZero, Copyleaks, and Originality.ai are separate competing products.
Can Turnitin detect Claude, Gemini, or newer models?
Turnitin's first model version was trained on GPT-3, GPT-3.5, and variants such as ChatGPT, and its 2023 whitepaper lists broadening to other LLMs as future work. Turnitin has updated the model since, but it does not publish a per-model detection table.
Why does Turnitin need 300 words?
The document score is an average of sentence scores. With too few sentences the average swings wildly, so Turnitin sets a 300-word floor to keep predictions stable.
Does the AI score appear in the plagiarism report?
It appears alongside it. The similarity score (text matching other sources) and the AI writing score are two separate numbers produced by two different systems, shown in the same Turnitin interface.
The same question, asked other ways
- What AI checker does turnitin use?1,000/mo