False positives · a postgraduate thesis or dissertation
Why your thesis reads like AI
If a methods chapter or literature review got flagged, that is not a coincidence. Those chapters are the most formulaic sections of any thesis by design, and formula is exactly what a classifier is built to notice.
A thesis methods chapter and a literature review follow a fixed template for good reason: examiners expect a specific structure, a hedged register, and dense citation. That same formula is what a classifier reads as low-variation, machine-shaped text: close sentence lengths, a narrow set of formal transitions, consistent passive constructions. The chapters most likely to get flagged are usually the ones written most correctly by disciplinary standards. Rewriting them to sound more varied would not make the thesis more honest. It would make it worse scholarship.
Why this happens
Every research discipline trains its writers toward a narrow, shared register for methods and literature review sections: past tense, passive constructions, hedged claims, a fixed sequence of moves such as design, then sample, then procedure, then analysis, and transitions drawn from a small formal set. Supervisors teach this on purpose, because examiners read hundreds of theses and rely on the convention to locate information quickly. The convention produces exactly the properties a classifier is trained to flag: low sentence-to-sentence variation, predictable transitions, and a hedge-heavy register that reads as generic caution rather than a specific voice. A student who followed the convention well will often score as more machine-like than one who broke it, because the classifier cannot distinguish disciplinary competence from statistical flatness.
Where honest writing and machine writing genuinely overlap
These are real patterns, not coincidences. Each one is something a careful human writer in this situation has good reason to do, and something a model does for entirely different reasons.
- The fixed methods sequence
- Methods sections are taught to follow the same order across a discipline: design, participants, procedure, analysis. The predictability of that sequence is a feature for examiners and a red flag for a classifier trained on generated text.
- The hedge stack
- Academic caution language, 'may suggest,' 'appears to indicate,' 'could be attributed to,' is required register in most fields. It also flattens the sentence-to-sentence variation a classifier uses to separate human from machine writing.
- The multi-citation summary sentence
- A single sentence citing several sources with a generic reporting verb, 'X argues, Y notes, Z suggests,' is standard literature-review practice. It reads structurally close to the sourceless vague-attribution pattern a classifier is trained to catch, even though every claim here is properly cited.
- The passive-voice convention
- 'Participants were recruited' and 'data were collected' are the required register in many empirical fields, written to keep focus on the method rather than the researcher. The same construction reads as flat and agentless to a tool built to notice a missing first-person voice.
What actually helps
Your reference manager and word processor already keep the evidence: dated versions, tracked changes, and comments from supervision meetings. An annotated bibliography or research diary, if you kept one, shows months of reading behind a literature review that might otherwise look like it appeared all at once. Bring the specific chapter to your supervisor before it goes anywhere near a misconduct process. Most supervisors have seen a formulaic methods section flagged before and can speak to whether the writing matches your usual style and your supervision history. Ask which tool was used and at what threshold the flag was set, since these thresholds vary widely and a borderline score is a different conversation than a high-confidence one.
What the flattening looks like
On the left, honest writing with the specifics drained out of it, which is the shape that reads as machine-written. On the right, the same content with the writer's own detail restored. The voice is The Operator (RCOEN): states the methodological decision once, names its source, and moves to the next section.
This study adopts a qualitative research design to explore the phenomenon in depth. Previous studies have shown that semi-structured interviews are effective for capturing participant experience, and this approach was therefore selected for the present research. The methodology section outlines the sampling strategy, data collection procedures, and analytic framework used to address the research questions. Ethical approval was obtained prior to data collection, and all participants provided informed consent.
This study uses a qualitative design because depth of understanding mattered more than breadth. The method is semi-structured interviews—they have a track record of catching what participants actually mean, not just what they are willing to say in a checkbox. The methodology section covers the sampling, the procedure, and the analytic frame that answers the research questions. Ethics clearance came first. Participants signed informed consent. That is the line between data and extraction, and the study does not cross it.
Steps worth taking
- Pull your version history. Your reference manager and word processor keep dated drafts and tracked changes. That trail is the plainest record of how the chapter was actually written.
- Gather your research diary or annotated bibliography. If you kept notes while reading, they show the work behind a literature review that might otherwise look assembled all at once.
- Go to your supervisor first. Bring the flagged chapter to the person who has read your writing across the whole project, before it reaches a formal process. They can speak to whether it matches your usual work.
- Ask about the specific tool and threshold. Detection thresholds vary widely between tools and departments. A score just over a threshold is a different situation than a high-confidence flag, and it is worth knowing which one you are dealing with.
Questions
Why is my paper getting flagged for AI when I wrote it?
Academic writing, especially methods and literature review sections, follows a narrow, disciplinary register on purpose. That register produces the low sentence variation and formal transitions a classifier is trained to associate with generated text, regardless of who wrote it.
Why is my dissertation being flagged as AI?
The most disciplined, correctly formatted chapters are often the most formulaic, and formula is what these tools measure. A flag on a methods chapter is not unusual and is not, on its own, evidence of anything.
How much AI is acceptable in a thesis?
That is a question for your department's policy, not for a detector score. Policies differ by institution and by field, so check yours directly rather than infer it from a tool's output.
Can academic writing be flagged as AI?
Yes, routinely. The formal, hedge-heavy, citation-dense register that most fields require sits close to the statistical profile these classifiers are built to catch, which is a known limitation of the tools rather than a property of your writing.
underslop edits drafts for voice. It does not check, score, or contest anything, and nothing on this page will change a result you have already been given. What it can do is show you which patterns in a piece of writing read as machine-written, using the same open lint the rest of the site runs on.