How Turnitin AI Detection Works in 2026 (The Technical Truth)

The public release of ChatGPT changed everything for academic integrity. And so, the essay, once the gold standard of student assessments, is now also being besieged. In response, Turnitin, the global heavyweight of plagiarism detection, launches an AI writing detection feature.

But what I want to know is: how does it know the difference between thoughts in a human head versus an algorithm designed by machines? If a student types out an essay - can Turnitin tell whether that was written by them or just by prompting a bot to write it?

Here we go on a deep dive into the maths, mechanics and controversy surrounding Turnitin's AI detector.

1. The Core Concept: An AI Catching an AI

Let me explain how Turnitin works by explaining how Large Language Models (LLMs), such as GPT-4, work. LLMs are essentially "next-word prediction machines" that statistically define the relationship between words. An AI writing something will typically choose the most statistically probable next word in any given sequence.

The idea behind Turnitin's detector is very straightforward: human writing is messy, it's idiosyncratic, sometimes clumsy. The way AI writes is going to be pretty much consistent, highly probable and the kind that scores average grades.

Turnitin's detector is a model that is based on the transformer architecture. More specifically, it uses the Bidirectional Encoder Representation from Transfomers (BERT) architecture. It is trained on a very large data set. It contains two types of writing:.

Human writing: Sourced from years of student submissions stored in Turnitin’s database.
AI writing: Generated by various models (GPT-3, GPT-3.5, GPT-4, etc.).

The detector then compares the two data sets to learn the statistical "fingerprints" that a human would produce as opposed to a machine.

2. Under the Hood: The Detection Process

When you submit your paper to Turnitin, the AI-detection process runs separately to the regular plagiarism check. Here is how it works step by step:

Step 1: Segmentation

The algorithm doesn't read the paper as a single document. Instead, it breaks it down into smaller chunks (like sentences or very short paragraphs, roughly 100-200 words).

Step 2: Check "Perplexity" and "Burstiness"

Turnitin's algorithm uses a proprietary method but all AI-detection systems look at two key things:

Perplexity: This is how surprised the model is by what you've written. If it has low perplexity, it means the words you've used are just what the model expected them to be. That suggests AI writing. If it has high perplexity, that means the words you've used are unusual or unexpected and that suggests a human wrote it.
Burstiness: This is how varied your sentences are. Humans are "bursty" - we write long complicated sentences and then a short simple one. But AI models are more consistent and rhythmic.

Step 3: Scoring Segments

The model then gives each segment a score out of 1, showing how likely it is to have been written by AI. If you get a score of 0.9 for a sentence, that means the model is 90% sure that sentence was written by AI.

Step 4: Aggregation

The system then works out an overall score by averaging all the segment scores together. What that score doesn't show is how confident the system is (like "we're 20% sure this is AI"). What it does show is what part of your document it thinks was written by AI (like "20% of this text was written by AI").

3. Interpreting the Results

This is important for both educators as well as students alike.

The Blue Percentage: It's usually separate from that famous "Red" similarity report (for plagiarism). That's because the AI score is unlike the similarity report.
The Threshold: Turnitin has built-in some logic to suppress the score if less than 20% of the text is generated by the AI, to avoid false positives (when the algorithm accuses a student over just a couple of coincidental sentences).
Highlights: It'll also highlight sentences it thinks were generated by an AI. Keep an eye out for whole paragraphs being highlighted, while skipping citations or quotes.

4. The Controversy: False Positives and Bias

Turnitin claim that they have a very high accuracy rate - stating less than 1% false positives on documents containing more than 20% AI generated content. Less than 1% can be huge if you apply it to millions of students. For a deeper look at accuracy, see Is Turnitin AI Detector Accurate?.

The "False Positive" Problem

When you write a text by yourself but AI detector marks it as a work generated by AI, it is called a false positive. A false positive is most likely to happen with:

Formulaic Writing: Technical writing, laboratory reports or legal definitions that require strict format can be flagged as false positives due to their lack of randomness.
Repetitive Phrasing: Students who lack a wide vocabulary and who rephrase the same sentences may accidentally produce a text statistically similar to the text generated by a computer.

The ESL/ELL Bias

The 2023 Stanford study found that AI detectors are much more likely to flag writing by non-native English speakers as generated by AI.

Why? Non-native speakers often tend to use less sophisticated vocabulary and straightforward grammar to avoid making mistakes. That makes their writing look like it was produced by a machine.

5. Can Turnitin Detect Humanizers and Paraphrasers?

And this is the ongoing "arms race."

If a student uses ChatGPT to generate some text and then changes the order of the sentences and the synonyms using one of these paraphrasing tools (for example, Quillbot), the ability of Turnitin to detect that still pretty high.

If a student takes text produced by an AI and manually rewrites 30 percent of that text - think of it as introducing human chaos and syntax errors - Turnitin's detection score usually degrades but it is still able to detect it.

Using really smart humanizers like RewriteAI can reduce Turnitin's detection score significantly. See our Best AI Humanizer guide for a full comparison.

6. Best Practices for Educators and Administrators

What does this mean for how you use these results?

Treat the AI score like a signal, rather than evidence: if a student gets a high score, have a conversation about why, instead of automatically giving them a failing grade
Look for the "tell". Any actual AI hallucination is more convincing than the detector score itself (and that's what this means). Does the paper say the students cited a book that doesn't exist? Did they say some historical event occurred on the wrong day? This is actually what we call "hallucination", when AI makes up facts to support what the student wants to argue!
What Turnitin detects, though, is whether or not the student wrote this in their own words!
Check the version history on the Google Doc If you get back to your teacher a flagged student and then ask to see the version history on that Google doc! A real writer writes all hour(s), adding and subtracting as necessary! But an AI-pasting student drops the WHOLE THING into Google Docs at once! If the score is only 25%!
Take another look at what actually gets highlighted. This could be a false positive. The score might just be picking up on some fluff. Context still matters.

Conclusion

That being said, Turnitin's AI detector is a sophisticated piece of technology that uses statistical probability to make a pretty educated guess about whether a text originated here versus there. But it's not magic - it detects predictability, not intelligence. As things continue to evolve, the lines around what constitutes human assisted writing vs. AI generated writing get murkier by the minute. For now though, Turnitin gives us sort of a weather forecast for academic integrity - it's usually right but don't ruin a students life without at least checking the forecast yourself!