AI detection and a thesis or dissertation

A long document written over years, edited by several people, is a specific and unusually awkward case for a classifier.

A thesis differs from every other document a detector sees. It is very long, it was written in passes over a long stretch, several people have edited it, and parts of it were probably published elsewhere first. Each of those changes what a score means.

Length helps, and then it stops helping

Detectors are far more reliable on long text than short. Turnitin's own figures make the point: under 1% of documents against around 4% of sentences. A 200-page thesis gives a classifier more evidence than anything else it will read.

That is the good news, and it comes with a catch. A single percentage over 60,000 words is an average across chapters that were written years apart in different states of mind, some of them heavily revised and some barely touched. The document number is stable and it describes nothing in particular. See documents that are part human, part AI.

The number worth having is per chapter, or per section.

Which chapters carry the risk

The literature review. Long, formulaic, densely hedged, structurally repetitive by necessity. It also tends to be written early and revised most, which is the worst combination for a score.

Methods. Formulaic by design, often reusing phrasing from your group's protocols or your own earlier papers, deliberately narrow vocabulary.

The abstract. Short and templated, so it lands in the least reliable regime any detector has.

Discussion and conclusions. These usually score lowest, because that is where the writing is most specific and least conventional.

That ordering is a property of the genre, not of how any given thesis was produced. A high score on a literature review is close to expected. Full mechanism in AI detection in academic papers.

Two things that raise the score with no model involved

Editing passes. Supervisor comments, a proofreader, a language editing service, five rounds of your own revision. Every one of them smooths prose, and smoothing removes exactly the variation detectors read as human. Measured on our corpus, human academic writing has a within-document sentence length standard deviation of 8.7 words against 6.4 for generated text, and each editing pass moves a document toward the lower figure. See does Grammarly get flagged as AI.

Writing in English as a second language. Repeatedly measured as an elevated false positive population, for the structural reason that learned English applies its rules more consistently than acquired English does. See why non-native English writing gets flagged.

A well-supervised thesis by a careful second-language writer is close to the worst case a detector has, and none of that says anything about how it was written.

Reusing your own earlier text

Chapters built from your published papers are normal in a thesis. Worth knowing that this affects the two scores on a report differently.

The similarity score will pick up your own published text as matching text, because it matches. What follows from that is a matter for your institution's own rules on prior publication, which vary and which we are not the right people to summarise.

The AI indicator is unaffected by whether text appeared before. It only reads how the writing reads. Published text that went through a journal's copy-editing has been smoothed more than your draft chapters, so those sections may well score higher than the rest.

Checking one before submission

Run whole chapters rather than the document or an extract. Chapter-level is the unit where a score is both reliable and interpretable.

Score the prose, not the .tex source, if you write in LaTeX. A detector reading your preamble and equation environments as prose returns a number about your markup. See why LaTeX breaks AI detectors.

Compare chapters against each other rather than against an absolute threshold. The interesting signal is a chapter that stands well above your own baseline, not a chapter above some number.