Checking a thesis for AI: what your university's detector will actually read

How much of a thesis a detector will read

A thesis, a dissertation and a final-year project are the same object as far as software is concerned: one long, chaptered document. The first thing any checker does with it is decide whether it is the right size. Turnitin publishes its answer in the file requirements for an AI Writing Report (read 26 September 2026):

The last row is the one thesis writers trip over. Turnitin’s detection FAQ says a gray indicator with no percentage means the indicator “is unable to process this submission”, and among the conditions it lists under that result is a file that exceeds 30,000 words. An oversized file is not scored on its first 30,000 words; it is not scored at all. So a 45,000-word thesis uploaded as one file should be expected to return nothing, rather than a partial reading.

The ceiling has moved before. The University of Bristol’s Digital Education Office told its staff on 17 July 2024 that Turnitin had raised the maximum from 15,000 to 30,000 words, and Turnitin’s release note of 16 July 2024 puts the new figure as “up to 30,000 words of qualifying text”. The file-requirements page simply says the file must not exceed 30,000 words. Those two wordings are not reconciled anywhere we could find, so if your document sits close to the line with a long reference list, plan on the stricter one.

Whether your department splits a long thesis across several submissions, or switches the AI indicator on for thesis submissions at all, is an institutional setting. Nobody outside your university can see it, and the graduate school office is the place to ask.

Per-check limits on other detectors

Self-checking tools publish their own ceilings, and they do not share a unit. Some count words and some count characters. We have left each figure in the unit its vendor uses rather than convert one into the other, because any conversion would be our estimate wearing their name.

All third-party figures were read 26 September 2026. GPTZero is missing because its pricing page showed no per-scan limit for any tier when we read it that day, and we do not quote a figure we could not see on the vendor’s own page. To compare a character limit with your own chapter, use the character count your word processor shows next to the word count; it is quicker and more honest than any rule of thumb. AI detection in PDFs covers the file-format side, including why a scanned thesis page carries no readable prose at all.

Why one chapter scores differently from another

Turnitin does not read everything in a thesis. Its guide to using the AI Writing Report defines what it calls qualifying text as “prose sentences contained in long-form writing format” and names the dissertation as an example. The same passage lists what the model does not reliably detect: poetry, scripts, code, and short-form writing such as bullet points, tables and annotated bibliographies. Set against the parts of a typical thesis, that gives a map:

For anyone reading a thesis score, the first consequence is that the percentage is taken over qualifying text only, and Turnitin’s FAQ warns that it “is not necessarily the percentage of the entire submission”. A data-heavy results chapter can therefore contribute very little prose to the calculation, while a discursive literature review contributes nearly all of its words.

The second consequence is about which prose gets misread. The same FAQ says false positives can include “content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas”. Read that list next to a thesis. A literature review that reports one source per paragraph is paraphrase by design. A methods chapter that follows the template your discipline expects has little structural variation by design. Our inference, and it is only an inference from Turnitin’s own description rather than a measured result, is that these are the chapters most exposed to a misleading reading. We have no per-chapter test data and are not claiming any.

For the mechanism behind that, see how AI detectors work; for the documented patterns in human writing that get flagged, see why human writing gets flagged. How to revise a literature-review chapter so it argues rather than lists is a writing question, and it has its own page: working on a dissertation chapter.

Length inside a chapter matters as well. Turnitin’s FAQ says that in documents of only a few hundred words the prediction is “mostly ‘all or nothing’”, because the model is scoring a single segment with nothing to overlap. An abstract or an acknowledgements page checked on its own is exactly that kind of document. Does text length affect AI detection? answers the general version of the question.

Thesis checks by country

Several of the rules below were written for similarity, which is a different measurement: a match against sources that exist, as opposed to a statistical guess about who produced the prose. Where an institution has written about AI and theses specifically, we quote it below. Where we could not find a primary source, we say so and leave the country out.

India

The UGC’s 2018 plagiarism regulations name the thesis directly. Institutions must use software “so as to ensure that documents such as thesis, dissertation, publications or any other such documents are free of plagiarism at the time of their submission”, and each student signs an undertaking that the document “has been duly checked through a Plagiarism detection tool approved by the HEI” (Compendium, p.514). Supervisors certify the work as plagiarism-free, Masters and research theses go to INFLIBNET’s Shodhganga repository within a month of the degree, and references, bibliography, contents, preface and acknowledgements are left out of the similarity check (p.515). All of this concerns similarity, not AI. The graded bands and what the regulations leave unsaid about AI are on our page on UGC rules. Which software reads your thesis depends on the degree: the national ShodhShuddhi licence covers PhD theses, their chapters and synopses, and research articles, and its revised guidelines rule out checks for undergraduate and postgraduate students, so a Masters dissertation or a final-year project goes through whatever tool your institution licenses for itself. Our DrillBit page quotes the scope rule in full, and your department office can say which tool applies and whether it produces an AI report.

Philippines

Philippine thesis guidebooks can fix the chapter structure in advance, which makes a single chapter a natural unit to check. One published example, Pacific InterContinental College’s thesis guidebook of January 2025, sets five chapters: The Problem and Its Background; Review of Related Literature and Studies; Methodology; Presentation, Analysis and Interpretation of Data; and Summary, Conclusions, and Recommendations. Going by the titles alone, chapters 1 and 2 are mostly prose, while chapter 4, which presents data, is the likeliest home for the short-form tables Turnitin does not read.

De La Salle and Ateneo de Manila have both written rules that reach thesis work. De La Salle University’s policy lets faculty use detectors such as Turnitin and GPTZero to flag submissions, but says they “cannot refuse submissions or impose grade penalties solely because of AI detector results”, and that this holds “regardless of any threshold value”. Its appendix adds that detectors are “mostly limited to English text” and that Filipino is not fully supported. Ateneo de Manila’s guidelines, approved on 18 November 2024, require intended use of generative AI to be disclosed in advance to capstone, thesis and dissertation advisers, and warn the community not to “depend fully on GenAI detection technologies” for high-stakes decisions.

Malaysia

Universiti Malaya publishes two documents in one compiled PDF, and together they are the most thesis-specific rules we read. Its Academic Policy on AI says AI use in “thesis writing or final research projects” must be approved by the supervisor or the relevant Responsibility Centre. Its Guidelines on the Ethical Use of AI (Version 2, Senate 20 November 2025) require students to make “a formal declaration of AI usage in producing their thesis, dissertation, or research report” on official forms, and say detection software “should not be used as conclusive evidence” for a misconduct charge.

Before you self-check, note two clauses. Section 7(5) of the guidelines says staff, supervisors, students and examiners may not upload a thesis, dissertation or research report “to any public AI platform until the assessment process is completed and finalised by the Senate”. The Academic Policy defines public AI platforms as systems “accessible to the public via public internet networks”, which “may include third-party services where the AI infrastructure is not hosted within the University”. On detectors specifically, the guidelines say staff, students, supervisors and examiners “may, with University approval, use any appropriate software to detect potential unauthorised AI use”. A separate caution against “uploading any ongoing assessment in full” is addressed to academic staff, supervisors and examiners; for students, the Section 7(5) bar above is the rule. An online AI checker, HumanFlow’s included, is a third-party service hosted outside the university. If you are at Universiti Malaya, get that approval before you paste a chapter into any of them.

United Kingdom

UK universities set their own rules on AI and dissertations; no national regulator requires AI detection, as our UK regulator page explains, and how UK institutions configure Turnitin’s indicator is covered in Turnitin AI detection at UK universities. One institution’s published practice is a useful reference point. The University of Bristol’s Digital Education Office wrote in November 2025 that Turnitin results “should always be reviewed by subject experts and used as supporting evidence, not as a replacement for academic judgement”, and warned that a model update could change scores between drafts submitted before it and final versions submitted after.

Checking your own thesis before submission

Start from the fact that no third-party score is a preview of Turnitin’s. It is a different model trained by a different company. Turnitin itself does not stand still either: its October 2025, February 2026 and May 2026 model releases each say they “will not retroactively update previously-generated AI writing reports”. So a chapter checked in March and the final thesis checked in September may not even have been read by the same version of Turnitin. What a self-check can do is point you at passages worth rereading.

1. Split at chapter boundaries. Cut where the chapters already break, not partway through an argument, and keep each piece well above 300 words.
2. Leave out what Turnitin leaves out. Reference lists, data tables and bulleted lists are not part of its AI reading. Removing them from your own check keeps the comparison closer to what the institutional report measures.
3. Write down what you ran. The tool, the date and the chapter version. A score with no date attached is hard to use as evidence of anything later.
4. Read the flagged passages, not the headline number. A conventional methods paragraph that reads as flat is a different finding from a discussion section you do not recognise as yours.
5. Keep the process record. Dated drafts, version history, notes from supervisor meetings and emailed feedback on chapters show how the thesis was built. How to show you wrote it explains which records carry the most weight.
6. Check the upload rule first. If your university restricts where a thesis may be uploaded before examination, get whatever approval it requires before you paste a chapter into an online checker. At Universiti Malaya that means University approval for any detection software, on top of the Section 7(5) bar on uploading a thesis to a public AI platform before the Senate finalises assessment.

If your institution requires AI assistance to be declared, as Ateneo and Universiti Malaya do for theses, declare it. A self-check result, whatever it says, does not change that obligation.

If a chapter is flagged

Turnitin’s own position is that it “does not make a determination of misconduct” and that its percentage “should not be used as the sole basis for action” (FAQ). De La Salle and Universiti Malaya say the same about detectors in their own rules. So a flag on a thesis chapter opens a conversation with your supervisor or panel; it does not settle one. Bring the drafts and meeting notes, and ask which parts of the chapter were flagged, since a reference-heavy or formulaic section is a different conversation from a discursive one. What to do if you are falsely accused covers the first meeting, and the appeal process covers what follows if it goes further.

Where HumanFlow fits

We sell an AI detector and a humanizer, so read this section knowing we have an interest. Our detector works in English only. A chapter in Filipino, Malay, Spanish or Arabic is outside what it can read, and Turnitin’s list of supported languages is wider than ours.

In a single scan the Free plan takes 1,500 words, and the ceiling rises through the paid plans to 30,000 on Team. Whether a chapter fits in one scan depends on its length and your plan, and a thesis longer than the top cap still goes in parts. The table above gives every tier, and the pricing page has the rest. Any tool whose per-check ceiling is above your chapter’s length fits a one-pass check better than a plan of ours that is not; for the tools that count characters, compare against your document’s character count rather than its words.

A HumanFlow score tells you how our model reads your prose. It cannot tell you what Turnitin will report, what your examiners will conclude, or whether your university allows the check in the first place.

Related

• Turnitin AI detection, in full: what its indicator measures and the published research that disputes it
• For students: self-checking shorter coursework
• For research papers: shorter papers and journal submissions drawn from thesis work
• All AI detection guides