Can Turnitin detect ChatGPT in 2026? What teachers see

Every semester the same question arrives from two directions. Students want to know whether Turnitin will catch a paragraph they drafted with ChatGPT. Teachers want to know whether the percentage on their screen is something they can act on. The honest answer to both is the same: Turnitin does flag likely AI writing, and Turnitin's own documentation is unusually clear that the flag is a starting point rather than a finding.

This post walks through what Turnitin's AI writing indicator is, what the instructor actually sees, what the student sees, the word and language limits, and what Turnitin itself says about false positives. Everything about Turnitin here comes from Turnitin's own help pages and blog, checked on 3 November 2026 and listed in the sources. Then we look at how Indian colleges typically reach Turnitin, and finish with practical advice for both sides of the desk.

Lacewing Technologies is an AI and software development company in Navi Mumbai. We build and operate our own products alongside client work, and one of those products is an AI text detector, so we have spent a fair amount of time reading how the big institutional tool behaves and where it is honest about its limits.

What is Turnitin's AI writing indicator?#

It is a feature inside the Similarity Report, not a separate product most students will ever log into. When a paper is submitted to a Turnitin account that has AI writing detection licensed and enabled, the text is scored by a separate model from the plagiarism matcher.

Turnitin describes the mechanism in its September 2024 blog for academic leaders. The submission is broken into overlapping segments of a few hundred words, roughly five to ten sentences each. Every sentence gets a score between 0 and 1, where 0 means the model thinks a human wrote it and 1 means it thinks the whole sentence was generated. The document-level percentage is built from the averages of those segment scores.

At the time of that post, Turnitin said the model was trained to detect output from GPT-3.5 and GPT-4, the models behind ChatGPT, and that it was working to expand coverage. Its current guides also describe detection of AI paraphrasing and "bypasser" text, meaning output run through a spinner or humanizer, for English submissions. In August 2026 Turnitin's product updates page announced a move to a single unified indicator, folding AI-paraphrased text into the same blue "AI-generated" category rather than showing it separately in purple. Older submissions have to be resubmitted to get the new view.

So, can it detect ChatGPT? Yes, that is literally what it was trained on. The more useful questions are how confident it is, how it shows that confidence, and what it refuses to score.

What does the instructor actually see?#

Three things, in order of how much attention they get.

The percentage. A single number in the Similarity Report sidebar, labelled as the share of qualifying text the model believes was AI-generated. This is the figure that gets screenshotted and argued about.

The highlighted text. Clicking through opens the AI Writing Report, where sentences the model scored as likely AI-written are highlighted. Turnitin's guide is specific that only "qualifying text" is highlighted: sentences within paragraphs that make up longer written work. Poetry, scripts, code, bullet points, tables and annotated bibliographies are called out as formats the model does not reliably detect.

The asterisk. If the model detects some AI writing but the amount falls below 20% of the document, the instructor does not see a number at all. The report shows an asterisk, written as *%, with no highlighted sentences. Turnitin's help centre explains this directly: below the 20% threshold, no numerical score is attributed and no text blocks are highlighted. The reason, given across its blog and guides, is that its testing found a higher incidence of false positives in the 1% to 19% range. Reports generated before 8 July 2024 may still show those low numbers, because the rule was applied to new reports.

There is one more thing the instructor sees that is easy to miss: Turnitin's own caveat. The guide for the AI Writing Report says the model can misidentify human-written, AI-generated and AI-paraphrased text, that further scrutiny and human judgment are needed, and that the report should not be the only evidence behind action against a student.

What does the student see?#

In most settings, nothing.

Turnitin's September 2024 post states plainly that "the AI writing score is visible only to Turnitin instructors and administrators". A student opens their Similarity Report, sees the familiar similarity percentage and coloured matches, and has no indication that a second model scored the same file.

Two nuances. Turnitin's current guide for the AI Writing Report describes an option for instructors to make the report available to students, in which case the student can open it from their Similarity Report and see the percentage and highlights. Separately, a Turnitin help centre article answering "why can't I see the AI writing detection report" says the report is not visible to students and suggests instructors download it and send the file if they want to share findings. Whether your college exposes it depends on the account configuration, which administrators control and can switch off for instructors too.

For an Indian student the practical implication is simple. You will usually not know your AI percentage unless a teacher raises it, so the only defence you can prepare in advance is evidence of your own process.

What does the indicator show, and what does it not show?#

Each row comes from Turnitin's file requirements page, its AI Writing Report guide or its blog, all linked below.

What are the minimum word counts and file rules?#

Turnitin's file requirements for an AI Writing Report are short and worth memorising if you set assignments:

• Minimum 300 words of prose text in a long-form format. Turnitin introduced this floor in May 2023 after finding that short submissions produced more false positives than it was comfortable with.
• Maximum 30,000 words of qualifying text. This ceiling was raised in July 2024.
• File types .docx, .pdf, .txt or .rtf, under 100 MB.
• Language must be one of the supported languages listed above.
• Licence the account must hold an active Turnitin Originality licence with AI writing enabled, otherwise the indicator simply does not appear.

A 250-word reflective paragraph, a slide deck, a spreadsheet, a problem set full of equations or a two-page Marathi essay will not get an AI score, or will get one Turnitin itself describes as less accurate. If you are a lecturer designing low-stakes writing tasks under 300 words, you are also designing tasks the indicator will not evaluate. That is not a bug; Turnitin built it that way to avoid guessing on thin evidence.

What does Turnitin say about false positives?#

More than most vendors, which is to its credit.

The headline figure. Turnitin states that the risk of a false positive from its indicator is less than 1% for a document with over 20% likely AI-generated content. That framing matters: the sub-1% figure is for whole documents that already crossed the threshold. It is not a claim about any single highlighted sentence, and it is not a claim about documents below 20%, which is exactly why those are hidden behind an asterisk.

On non-native English writers. Turnitin's India solutions page cites its own test of English language learner writing, reporting a false positive rate of 0.014 for ELL writers against 0.013 for native English writers on documents meeting the word count requirement, and concludes there is no statistically significant bias against non-native speakers. We would note two things. First, this is a vendor's own test of its own tool, which is what we can verify today. Second, a 1.4% rate across a university that submits tens of thousands of essays a year is still hundreds of students, so the rate being low does not make the process around it unimportant.

On how to use the score. This is the part worth quoting to any department head. Turnitin writes that the AI writing report "is not an absolute proof or disproof of AI writing", that it exists "to facilitate formative conversations with students", and that "an accusation should never be made without a respectful dialogue with the student in question". It recommends offering the benefit of the doubt, gathering "further, more definitive evidence" before invoking the integrity policy, and, where an investigation cannot reach a definitive conclusion, "moving forward without accusation or penalty".

If a college's internal policy treats the percentage as a verdict, that policy is stricter than the tool's maker says the tool can support.

How do Indian universities actually use Turnitin?#

Mostly through the library, and mostly for similarity rather than AI detection.

The national route is the Ministry of Education's ShodhShuddhi programme, run by INFLIBNET, which has provided plagiarism detection software to over 1,100 institutions since September 2019. INFLIBNET's own page records that this was Ouriginal (formerly Urkund, later owned by Turnitin) until the contract ended on 30 September 2023, and that since 1 October 2023 the programme supplies DrillBit-Extreme instead. So a state university relying solely on ShodhShuddhi is not running Turnitin at all today.

Institutions that do use Turnitin buy it directly. Delhi University's library system, for example, offers faculty both DrillBit through ShodhShuddhi and Turnitin accounts requested through the Samarth portal, with colleges asking the university librarian for Principal, IQAC and library accounts. That pattern, a central or library-held licence with a handful of named accounts per department, is typical of the Indian campuses we have spoken to, though we have not surveyed it systematically and you should check your own institution's library page.

Two consequences follow. Faculty access is often mediated by a librarian or IQAC coordinator rather than being embedded in every course, so AI scores may be generated on a thesis at submission time rather than on weekly assignments. And whether the AI writing indicator is even switched on depends on whether the institution licensed Turnitin Originality and whether the administrator left the AI setting enabled. Many Indian students worrying about an AI percentage are submitting to accounts that only produce a similarity percentage.

Should students self-check before submitting?#

If your institution permits it and your assignment rules allow AI assistance, running your own draft through a checker before submission can tell you whether your writing reads as formulaic. Keep expectations modest: no third-party tool runs Turnitin's model, and a low score elsewhere does not predict a low score in Turnitin. Options students in India commonly try, in no particular order of accuracy:

1. GPTZero positions itself as an education-focused checker; its homepage box accepts up to 10,000 characters without an account, and asks you to create a free account for longer scans.
2. Originality.ai is aimed at publishers, agencies and schools, and offers free scans with a limit of 100 credits per account; it also notes that text pasted into the free detector may be used for training.
3. Copyleaks offers a free AI detector check of up to 25,000 characters per scan and documents support for over 30 languages, including Hindi.
4. TextSight, which we build, scans up to 1,500 words free and returns a score with the passages that drove it; it is positioned as a signal for review, not a verdict, and does not claim to mirror Turnitin.
5. QuillBot's AI Detector allows free scans of up to 1,200 words, six per day, and recommends at least 80 words for a usable result.
6. HumanFlow, also ours, is primarily a humanizer that rewrites for rhythm and specificity, with a built-in detection score and 10,000 free detection words a month. Rewriting that way tends to lower detection scores because variation is what detectors read as human, but it is not a guarantee, and if your institution requires disclosure of AI use you must still disclose.

Pasting an unpublished thesis into any free web tool has its own risks. Read the data policy first, and never upload another person's work.

What should teachers do with the number?#

Treat it as a reason to look, not a conclusion. Turnitin's own guidance frames the score as a conversation starter. Read the highlighted passages. Quoted definitions, standard methodology boilerplate and heavily templated sections often account for a chunk of the highlights.

Ask for process evidence. Turnitin lists research notes, outlines, document version history and metadata, earlier drafts and feedback received as the kind of material a genuine author will usually have. Google Docs and Word version history are the cheapest evidence there is; make asking for it routine rather than accusatory.

Compare against known writing. Turnitin suggests collecting an in-class writing sample at the start of term, with the reason explained, so there is a baseline to compare against later.

Publish your policy. Tell students at the start of the course that the indicator is on, what it can and cannot show, which uses of AI are permitted, and how a flag will be handled. A process that cannot be explained to the student it judges is not a process.

When in doubt, do not penalise. Turnitin's words, not ours: if the investigation cannot reach a definitive conclusion, move forward without accusation or penalty.

What should students do?#

Keep your drafts. Write in a tool that records history, keep your notes and outline, and do not delete the messy early versions. If a conversation ever happens, that history is what settles it.

Write past 300 words of real prose when you can. Not to game anything, but because the indicator is less reliable on short text and because a longer piece in your own voice gives both you and your teacher more to work with.

Disclose where required. If your college or your assignment asks you to declare AI assistance, declare it. A declared use of a grammar tool is a footnote; an undeclared one discovered later becomes an integrity case regardless of what the percentage said.

Do not treat any external score as a pass. A self-check can tell you your writing is formulaic enough to be flagged. It cannot tell you Turnitin will agree.

Turnitin can detect ChatGPT, within the limits it documents: 300 words or more, a supported language, long-form prose, and a result it only states as a number once it clears 20%. The tool's maker says the number is a signal for a conversation. The most useful thing an Indian college can do in 2026 is write that sentence into its own integrity policy.