If you are wondering how Turnitin detects AI, the short answer is that it looks for patterns across the submitted writing. It does not search the internet for a matching ChatGPT response, inspect your browser history, or ask an AI chatbot whether it wrote the paper. Instead, Turnitin uses a trained AI writing detection model to estimate whether passages of qualifying prose resemble human writing or text produced by a large language model.
That estimate is useful, but it is not proof. Turnitin itself says its model can misidentify human-written, AI-generated, and AI-paraphrased text. An AI Writing Report should therefore be treated as one piece of evidence alongside drafts, notes, citations, version history, and an instructor's judgment.
This guide explains the process in plain English, separates AI detection from plagiarism checking, and shows how to run a private pre-submission review with Turntin.app's free AI detector.
How Does Turnitin Detect AI Writing?
Turnitin describes a multi-stage detection process. When an eligible paper is submitted, the system extracts prose sentences and divides them into overlapping segments. Its detection model then classifies those segments according to how likely they are to be human-written or AI-generated.
Because the segments overlap, a sentence may contribute to more than one prediction. The model pools those results into sentence-level scores and then combines the sentence scores to produce an overall AI writing percentage for the document's qualifying text.
In simplified form, the workflow looks like this:
- Turnitin identifies long-form prose that qualifies for analysis.
- It splits the prose into overlapping groups of sentences.
- A trained classifier evaluates each segment.
- Overlapping predictions are combined into sentence-level estimates.
- Sentence results are aggregated into the document's AI writing score.
- The report highlights qualifying passages that the model considers likely AI-generated or AI-altered.
Turnitin describes its detector as using a transformer-based deep-learning architecture. The exact production model and all its features are proprietary, so claims that Turnitin relies on one simple measurement—such as perplexity alone—are incomplete. The result comes from a learned combination of language patterns rather than a single keyword, phrase, or grammar rule.
What Patterns Can an AI Detector Recognize?
Turnitin does not publish a checklist that lets someone reproduce or predict every decision made by its model. More generally, AI writing classifiers learn differences between examples of human and machine-generated writing. Those differences can include relationships involving:
- word and phrase predictability
- sentence construction and variation
- repeated or formulaic transitions
- consistency of tone and syntax
- patterns across neighboring sentences
- broader statistical relationships within long-form prose
None of these features proves AI use by itself. A careful student may write consistent, formal prose, while AI-generated text may be extensively edited. That is why the system evaluates patterns across segments and why its result remains probabilistic.
It is also why simplistic advice about changing a few words or making sentences less polished is unreliable. The detector is not merely counting sophisticated words or checking whether every sentence has the same length.
Turnitin AI Detection Is Not the Similarity Score
This distinction is essential. Turnitin's Similarity Report compares submitted writing with material in its databases, such as websites, publications, and previously submitted papers. It reports matching or similar text; it does not determine by itself whether plagiarism occurred.
The AI Writing Report answers a different question. It estimates how much of the qualifying prose may have been generated by an LLM or generated by AI and then modified with an AI paraphraser, word spinner, or bypasser tool.
A paper can therefore have:
- a low similarity score and a high AI writing score
- a high similarity score and a low AI writing score
- low scores in both reports
- matches caused by correctly quoted and cited material
The two percentages have different denominators and should never be added together or interpreted as the same measurement. For a detailed explanation, see our guide to Turnitin AI detection and similarity reports.
What Text Does Turnitin Analyze?
According to Turnitin's current guidance, the AI Writing Report focuses on qualifying text: prose sentences in a long-form format such as an essay, article, or dissertation.
Turnitin says the model does not reliably analyze some other formats, including:
- poetry
- scripts
- source code
- bullet points
- tables
- annotated bibliographies
- other short or unconventional formats
This matters when interpreting the displayed percentage. If a document contains prose plus tables, references, lists, or other excluded material, the AI percentage refers to the qualifying prose—not necessarily every word in the file.
Current Turnitin file requirements say an AI Writing Report needs at least 300 words of long-form prose, must not exceed 30,000 words or 100 MB, and must use a supported file type and language. Availability also depends on the institution's license and settings.
What Does the Turnitin AI Percentage Mean?
The displayed percentage represents the share of qualifying text that Turnitin's model identifies as likely AI-generated or AI-generated and later altered by certain AI rewriting tools. It is not a probability that the student cheated, and it is not a confidence score for the entire paper.
For example, a 35% result does not mean there is a 35% chance the writer used AI. It means the model classified approximately 35% of the eligible prose as likely AI-written or AI-altered under its current detection system.
Turnitin currently treats low results cautiously. When the detected amount is above 0% but below 20%, the report shows *% rather than an exact percentage and does not display highlights. Turnitin explains that false positives occur more often in this range, so hiding the precise score helps reduce overinterpretation.
At 20% or above, the report can display a percentage and highlighted passages. Even then, the highlights are prompts for review, not a misconduct verdict.
Can Turnitin Detect Paraphrased AI Text?
Turnitin says its English-language detector can identify qualifying text that appears to have been generated by AI and then modified by an AI paraphraser, word spinner, or bypasser. Capabilities can differ by supported language and change as Turnitin updates its models.
This does not mean every paraphrased passage will be detected. Rewriting exists on a spectrum: correcting a typo, accepting a grammar suggestion, manually revising a paragraph, and asking an AI tool to rewrite an entire section are not identical activities. Detection accuracy also varies with text length, model, language, editing depth, and writing style.
If you use writing assistance, follow your course policy and keep evidence of your process. Draft history and research notes often provide more meaningful context than trying to infer authorship from one percentage.
Can Turnitin Detect ChatGPT, Gemini, or Claude?
Turnitin's detector is designed to identify patterns associated with writing from large language models rather than identify one chatbot by name with certainty. It may flag prose generated by tools such as ChatGPT, Gemini, or Claude when that prose resembles examples recognized by the current model.
However, a report should not be read as “Turnitin proved this paragraph came from ChatGPT.” Different models, prompts, languages, revisions, and mixed human-AI workflows can produce different outcomes. If this is your main concern, read Can Turnitin Detect ChatGPT?.
How Accurate Is Turnitin AI Detection?
No AI detector is perfectly accurate. There are two important error types:
- False positive: human writing is classified as likely AI-generated.
- False negative: AI-generated writing is classified as human.
Turnitin says its false-positive rate is below 1% for documents in which more than 20% of qualifying text is identified as AI-written, based on its testing. That figure describes a particular test condition; it is not a promise that every individual highlight or report is correct.
Turnitin's own guidance states that the model may make mistakes and should not be the sole basis for adverse action against a student. A responsible review considers the assignment, the writer's previous work, source use, drafts, document history, and an opportunity for the student to explain their process.
Why Human-Written Work Can Be Flagged
Human prose can share characteristics with AI output. Formulaic assignments, heavily standardized language, repeated sentence structures, generic summaries, or extensive editing can sometimes create statistical patterns a detector associates with generated writing.
A flag does not automatically mean the student did something wrong. If your own writing is questioned, preserve the submitted version and gather:
- outlines and handwritten notes
- source annotations and citation records
- earlier drafts
- Google Docs or Microsoft Word version history
- assignment-specific research
- a clear timeline of how the paper developed
Do not immediately replace the paper or erase its history. Process evidence is often the clearest way to support genuine authorship. Our guide on what to do if Turnitin says your work is AI-generated provides a fuller response plan.
Check Your Draft Free with Turntin.app
You do not always have access to your institution's Turnitin AI Writing Report before submission. Turntin.app gives students and writers a free way to review a draft for AI-like patterns first.
Our AI detector is designed to provide accurate, practical feedback without pretending that any detector can deliver a perfect verdict. Paste your text, run the check, and use the result to identify passages that deserve closer review. The goal is to strengthen your own reasoning, add specific evidence, and make sure the final work honestly reflects your voice—not to disguise prohibited AI use.
Turntin.app is also built around data safety:
- submitted text is processed only to return the requested analysis
- text is not retained after the check is completed
- submissions are not used to train machine-learning models
- submitted writing is not sold or shared with third parties
You can review the details in our Privacy Policy. Turntin.app is an independent service and is not affiliated with or endorsed by Turnitin.
For a quick check, use the free detector at Turntin.app. If you want an unrestricted Pro experience for longer or repeated workflows, visit AceEssay Pro AI Checker.
A Responsible Pre-Submission Workflow
An AI score becomes more useful when it supports good writing habits. Try this workflow:
- Write from your own outline, research, and notes.
- Cite borrowed ideas and quotations accurately.
- Keep drafts and version history as you work.
- Run the complete draft through Turntin.app.
- Read flagged passages in context instead of chasing a target score.
- Replace generic claims with your own analysis and specific evidence.
- Confirm that every edit complies with your institution's AI policy.
- Proofread the final version and save your writing-process evidence.
Never assume that a low detector score guarantees acceptance or that a high score proves misconduct. AI detection should support critical review, not replace it.
Frequently Asked Questions
How does Turnitin know if you used AI?
Turnitin cannot directly know who or what wrote a passage. Its model analyzes qualifying prose in overlapping segments, assigns likelihood estimates, combines sentence-level results, and reports text it classifies as likely AI-generated or AI-altered.
Does Turnitin compare a paper directly with ChatGPT?
Not in the way a similarity checker compares wording with a stored source. AI detection classifies linguistic patterns associated with human and machine-generated writing. It does not need to find an identical ChatGPT response online.
Is the AI score the same as the similarity score?
No. The Similarity Report measures textual overlap with sources in Turnitin's databases. The AI Writing Report estimates how much qualifying prose resembles AI-generated or AI-altered writing. They are independent results.
What does *% mean in Turnitin?
It means Turnitin detected a result above 0% but below 20%. Because results in this range have a higher incidence of false positives, Turnitin suppresses the exact percentage and highlights.
Can a human-written essay be flagged as AI?
Yes. False positives are possible, which is why a report must be evaluated with drafts, notes, version history, writing samples, and human judgment.
Is Turntin.app free and safe?
Yes. Turntin.app offers a free AI detection check and processes submitted text only to produce the requested result. According to our privacy policy, the text is not retained or used to train models. For an unrestricted Pro workflow, use AceEssay Pro.
Final Answer: How Does Turnitin Detect AI?
Turnitin detects AI by applying a trained classification model to eligible long-form prose. It divides sentences into overlapping segments, estimates whether those segments resemble human or AI-generated writing, pools the predictions, and calculates an overall percentage for qualifying text. That AI result is separate from the Similarity Score and should always be interpreted with human judgment.
If you want to review your work before an official submission, start with the free, privacy-focused detector at Turntin.app. Use the result as feedback, preserve evidence of your writing process, and follow your institution's rules on AI assistance.