Blog

· By Thesisroom

Updated

How Universities Detect AI Writing in Theses (And Why the Viva Is Better)

How do universities detect AI writing? Turnitin, GPTZero, and Copyleaks run different checks with different accuracy. Here is what the tools do, what they miss, and why the viva is the sharpest test.

A white desk seen from above with a keyboard, two rose gold pens, round glasses and a lined notepad.

The four detection layers universities actually use

How do universities detect AI writing in theses is a question with a specific answer. Four detection layers operate at most institutions, and each catches a different subset of AI-generated content.

Layer 1: automated AI detectors. Turnitin added AI detection to its plagiarism-check product in 2023. Turnitin’s own numbers report a document-level false positive rate below 1% at the threshold where a document is flagged as containing 20% or more AI-generated content, rising to around 4% at the sentence level, with a miss rate around 15% (Turnitin, 2026). GPTZero self-reports 99.3% accuracy and a 0.24% false positive rate, though independent tests show much lower real-world accuracy, sometimes around 63% (GPTZero, 2026; third-party comparisons).

Layer 2: plagiarism cross-check. Turnitin also runs standard similarity checks against its database of academic and web sources. AI-generated content often does not match existing sources directly, but paraphrases of source material still surface as similarity matches.

Layer 3: citation verification. Examiners and integrity officers now cross-check citations against Crossref, OpenAlex, or the DOI system. Fabricated citations from AI-assisted drafts fail these checks reliably.

Layer 4: the viva. The oral examination remains the sharpest test of whether the writing is the candidate’s. An examiner asking specific questions on the methodology, findings, and citations catches AI-generated content that no automated tool reliably flags.

Why the automated detectors have limits

AI detection tools are not reliable enough for universities to rely on them alone. Turnitin’s document-level threshold requires at least 20% AI content before flagging, which means partial use often goes undetected. GPTZero’s independent-test accuracy is much lower than its self-reported figure. Copyleaks and Originality.ai report false positive rates around 5% to 9% in third-party testing.

Detection accuracy is also unevenly distributed. Detectors show elevated false positive rates against non-native English writers, whose prose patterns can resemble AI output more closely (Turnitin, 2026 documentation; independent research). A detector-first policy risks flagging authentic writing from international candidates while missing polished AI content from native English writers.

This is why detector output is best treated as a signal for further review, not as a verdict. A high AI-content flag triggers investigation; it does not by itself decide misconduct.

The false-positive problem

The reliability limits above create a specific risk for candidates who did not use AI. A false-positive flag from Turnitin or GPTZero can trigger an investigation that takes weeks or months to resolve. The best defence is a paper trail:

  • Drafts saved at intervals (via version control or dated backups) show the writing developing over time
  • Reference management records show sources saved and cited on specific dates
  • Research notes and supervisor emails show the intellectual work progressing under human effort

Candidates who can produce this trail resolve false-positive flags faster than candidates who cannot. The Thesisroom method supports the reference-management part of this: sources verified as they were saved, chapters reviewed as they were drafted, and rehearsal sessions logged as they ran, each with dates.

The viva is the sharpest detector

Mullins and Kiley (2002), interviewing 30 experienced examiners, found that panels form early impressions from structure, references, and the opening chapter. An examiner asking specific questions in the viva catches AI-generated content faster and more reliably than any automated tool, because AI-generated arguments do not hold up under specific questioning from someone who knows the field.

The specific tests that expose AI content in the viva:

  • Asking about a specific claim in a specific paragraph and requesting the reasoning
  • Asking about a cited paper and requesting the argument the paper actually makes
  • Asking about a methodological choice and the alternatives considered
  • Asking about a finding and the primary data behind it

A candidate whose writing is their own answers these fluently. A candidate whose chapters were partly generated stumbles at the fourth or fifth follow-up, because the intellectual work was not done.

Preparation on the honest side

The safest preparation is straightforward: do the intellectual work yourself, use AI only for support (concept explanation, editing, brainstorming), and verify every citation before it enters the thesis. This aligns with what current university policies expect and with what the viva actually tests.

Verifying citations at scale is the specific job Thesisroom’s citation check runs, drawing on nine registries. Reviewing chapters for structural weaknesses is what Chapter Audit covers. Rehearsing anchored viva questions is what Defense Practice generates. None of the modules generates content for submission.

What Thesisroom does with this

Thesisroom is a checks-the-work tool, not a does-the-work tool. It verifies references, reviews chapters, and generates rehearsal questions from your own writing. The integrity policy sets this position formally.

For candidates worried about false-positive AI detection, the tool provides the specific paper trail (verified sources with dates, reviewed chapters with timestamps) that resolves false-positive flags quickly.

Frequently asked questions

How accurate is Turnitin at detecting AI writing?

Turnitin reports a document-level false positive rate under 1% at the threshold flagging documents with 20% or more AI content, rising to around 4% at the sentence level. Miss rate around 15%. Independent tests show similar order-of-magnitude accuracy. No detector is reliable enough to be a sole verdict.

Can universities detect ChatGPT specifically?

Not by name, but stylistic markers of generated text are detectable at meaningful rates. Turnitin, GPTZero, Copyleaks, and Originality.ai each attempt this with varying accuracy. All produce false positives, particularly for non-native English writers. Detector output is usually a trigger for further review, not a verdict.

What happens if I am falsely accused of using AI?

Most institutions have a defined procedure. You are notified of the flag, given a chance to respond, and can present a paper trail showing the writing developed under your own effort (drafts, research notes, supervisor emails, reference-management records). Resolution takes weeks to months. Candidates with a strong paper trail resolve faster.

Do detectors flag Grammarly-edited writing as AI?

Rarely. Grammarly corrects existing prose rather than generating new content, and the residue of Grammarly editing does not usually trigger AI-content flags. Some detectors are more sensitive than others; if you use Grammarly heavily and worry about false positives, keep drafts that show the original writing.

Does the viva actually detect AI use?

The viva is one of the most reliable detectors, because AI-generated arguments do not hold up under specific questioning from a reader who knows the field. Candidates whose chapters cannot be defended in specific detail are exposed regardless of what detection tools flagged in advance.

How does Thesisroom help with AI-detection concerns?

Thesisroom sits on the checks-the-work side of the AI line. It does not generate content, so nothing it produces enters your thesis. It verifies your existing citations, reviews your existing chapters, and generates rehearsal questions from your existing text. The paper trail from these sessions supports the honest use of AI in academic work.

One takeaway before you close this tab

Universities detect AI writing through four layers: automated detectors, plagiarism cross-check, citation verification, and the viva. The viva is the sharpest layer, because AI-generated content does not hold up under specific questioning. The safest position is doing the work yourself and keeping a paper trail. Thesisroom’s method supports the honest-use side.

References

GPTZero. (2026). AI detection accuracy and comparison to other tools.

Mullins, G., & Kiley, M. (2002). ‘It’s a PhD, not a Nobel Prize’. Studies in Higher Education, 27(4), 369-386.

Turnitin. (2026). Understanding false positives within AI writing detection capabilities.

viva · false positive · writing · content · universities