AI Checker Cries Wolf — Dissertation Goes on 'Spin'; or, Yes Virginia, AI Checkers for AI Writing That Actually Work Are a Myth

There is a particular kind of frustration that comes from being told your own writing is wrong — when it is not.

AI tools for AI-generated text detection do not detect AI use. They detect the features associated with AI-generated text:

• controlled sentence structure
• consistent register
• precise vocabulary
• low variance between paragraphs.

Alas, the problem is that those are exactly the features of well-written academic prose. The better your writing, the more likely a statistical checker is to flag it.

The University of Waterloo — a university that built its international reputation on computer science — discontinued Turnitin’s AI detection entirely in September 2025 after internal testing flagged human-written text as 100% AI-generated in more than one instance. It has not been reinstated. Waterloo is not an outlier. It is early.

You ran a check. The tool flagged a passage. You read it back. It looked fine — because it was fine. You rewrote it anyway, because the number felt authoritative — hey! The machine said so and your institution trusts the machine — or does it? However, you ran it again. It flagged something else. At some point in that process, you stopped trusting your own writing and started trusting the score.

That is the moment worth examining.

So what in heck is really going on?

First, AI — as we call it in everyday conversation — itself is not AI. It is machine learning (ML). It is not a self-teaching neural network, it is not like a human brain in architecture. It does not really reason. It is a humungous collection of algorithms weighing up the statistical probabilities of what ‘token’ will be asked for next — and much of what it is trained on is not correct in the first place. Then there is the issue of how much an AI session can hold in its memory — for anything serious? Often not at all enough.

The other day I fed a frontier AI model a passage from a historical novel draft. A character makes arguments about duty — quite serious ones to refute another character. The AI took nearly four minutes to think about the passage, and then? Incorrectly flagged faults in the argument — why? It did not check the actual history correctly. Due to memory and other limitations AI’s guess a LOT. They ‘assume’ as it were in human terms and we all know where assumptions go. Then, when confronted? It took (expensive) minutes more to re-examine its checking. It came up wrong again. Only after well over 10 minutes of token consumption and several cycles of being corrected did it finally admit that — yes — the character’s arguments were correct, at their face, and 100% so, and, additionally? Were period-correct. It had been wrong, over and over — and with the whole historical record in its greedy little electronic paws. Another takeaway and relevant to AI checkers of text? The AI struggled — and very badly — with genuine reasoning. Do I trust AI answers? No. Do I use AI? Yes. But with great caution especially with text – see our blog post The Ghost in the Machine

Back to AI checking tools. Again, they are statistical systems. They were trained on large collections of text and learned to recognize patterns associated with AI-generated writing — controlled sentence structures, consistent register, precise vocabulary, low variance between sentences. When your text matches those patterns closely enough, the tool flags it.

What the tool cannot do is check what your text means or where your claims come from. It cannot read the source you cited and compare it to what you wrote. It cannot assess whether your argument is sound or whether your interpretation is accurate. It measures only the statistical properties of the prose itself — the surface texture of the writing, not its content or origin.

This produces a result that surprises most candidates. The tool is not detecting AI. It is detecting the features associated with AI-generated text — and those features overlap significantly with the features of well-written academic prose.

Precise vocabulary. Consistent register. Clear, controlled sentence structure. Low variance between paragraphs. These are the qualities a good editor and a good doctoral supervisor will push you toward. They are also exactly the qualities an AI model produces when it generates fluent text — but textual fluency is not the story in a doctoral dissertation; research, insight, and genuine thinking being reported is.

The better your writing, the higher your risk of a false flag.

A candidate who writes in a loose, conversational, inconsistent style will score lower on AI detection than one who writes with the clarity and precision a dissertation requires. The detector is rewarding sloppiness and penalising competence — not because anyone designed it that way, but because the statistical overlap between good human writing and AI output is real and unavoidable.

This is why precision work — the kind that produces the clarity a committee needs — can inadvertently raise detection scores. Clarity is not a writing skill. It is whether you understand the thing you are saying. But when the thinking becomes clear, the prose becomes clean, and clean prose looks like AI prose to a statistical checker.

How long will this keep happening?

Indefinitely, in its current form. The problem is not a bug awaiting a patch. It is a structural feature of how statistical language systems work.

Also, AI algorithms change constantly. We often forget how new and frail the technology is. To you that means that these checkers - trained to detect AI-generated text based on their statistically-based understanding of that text – change their outputs continually. As these AI models constantly ‘improve’ and their output becomes more varied and naturalistic, checkers must continually retrain. As they retrain, their accuracy – such as it is – that is to say, the boundary between flagged and unflagged text shifts unpredictably. False positive rates will fluctuate but will not reach zero — because the checker is always working from surface statistics, never from meaning.

I am a computer scientist by background. My position, after years of following ML development, is that statistical pattern-matching cannot reliably distinguish between two things that look alike on the surface but differ in origin. It can only say: this resembles that. When ‘this’ is your best writing and ‘that’ is the pattern the tool learned to flag, you are caught in a statistical argument that has nothing to do with whether you wrote it or whether it is correct. At its very best construction? Possibly these checkers may point to areas for a human to look at. However? As we recommend on /chairai — it is faster, cheaper, and easier on all concerned for a Chair, Mentor, or Committee member to simply sit down the candidate and look at a few passages from their thesis and ask them to explicate their reasoning.

Bottom line? For now? Institutions are better off to skip expensive AI checkers and rely on far better human ones they already pay for — Chairs!

If you are facing an accusation

A detector score is not proof. It is (at best!) a flag. If your institution is treating a score as evidence, the tool’s documented false positive rate is relevant to your case.

The peer-reviewed research on what detector scores actually mean — and what they do not mean — is at /aidetection. That page also covers what to document and how to respond if you are formally accused. If the situation has escalated beyond a conversation with your chair, /advocacy has the triage framework.

What the check is actually good for

This is not (well…not totally…) an argument against using checking tools. It is an argument against treating their output as authoritative.

The most useful thing you can do with a checking tool is treat every flag as a question, not a verdict. Check the passage. Check the source. If the passage is correct, document that it is correct and why. Do not rewrite correct text because a pattern-matcher is uncomfortable with it.

Your writing is the record of your thinking. A tool that changes it without cause changes the record.

One of five posts on AI and doctoral writing. The Ghost in the Machine — a dated log of what these tools actually did to real text, passage by passage — is at /aifailures.