16 September 2026 · 8 min read
You wrote it. You read the sources, built the argument, revised it three times, and the report came back with a high AI writing score. This is now a common experience in Nepali institutions and it is genuinely frightening, because unlike a similarity match there is nothing to open and compare. This post explains why it happens, why the instinctive response makes it worse, and what actually answers the question.
The AI writing indicator and the similarity score are produced by different mechanisms and mean different things. Similarity points at specific text in a specific source — it is evidence, and it can be examined and argued with. The AI indicator is a statistical judgement about the character of your prose. There is no source behind it, nothing to open, and no passage-level attribution to check.
That is the whole difficulty. A student with a high similarity score can open the matched source and demonstrate that the passage is quoted and cited. A student with a high AI score has been handed a number with no underlying claim to rebut, which is why it feels unanswerable and why the response has to be evidential rather than textual.
It is also why a low similarity score offers no protection. The two figures are independent, and a document at 4% similarity referred on the AI indicator is now one of the most common ways a Nepali student first encounters an academic conduct query.
Detector false positives are not randomly distributed. They cluster on prose that is formal, structurally regular and vocabulary-conservative — which is a fair description of careful academic writing by someone working in English as a second or third language. Writers who learned English through formal instruction rather than immersion produce exactly the patterns these systems associate with generated text.
Heavy editing pushes in the same direction. A student who drafts, revises and tightens repeatedly ends up with smoother, more uniform prose than one who writes once and submits, and uniformity is part of what the indicator responds to. Doing the work more carefully can produce a higher score, which is the part students find hardest to accept.
Certain sections attract it structurally too. Methodology chapters, definitions, standard procedural descriptions and literature summaries are formulaic by design — the discipline requires that phrasing — and formulaic is what the detector is looking for. A flag concentrated in your methodology section is telling you something about the genre, not about you.
None of this is a conspiracy theory about the software. The limitation is documented by the vendors themselves, which is a useful thing to know if you end up explaining it to a department.
The instinct is to rewrite the flagged sections until the number falls. It rarely works and frequently backfires, because there is no editing operation that reliably lowers an AI score on honest work. You can spend a week rewriting and move the figure very little.
Running the text through a paraphrasing tool is worse. Those tools are themselves generative, so their output carries its own detectable signature — you can raise the score while also producing text that no longer says what you meant and that your supervisor will notice does not sound like you. This is the single most common way a student turns a recoverable situation into a much harder one.
There is also a practical risk in rewriting first: you destroy the version history that would have demonstrated how the text developed. The evidence you need is in the drafts, and overwriting them to chase a number removes the thing that would have answered the question.
The defence is the drafting trail, and it is built while you write rather than assembled when you are challenged. A student who can show three weeks of development is in a far stronger position than one who can only produce a finished file, whatever any indicator reports.
Approach your department with that evidence rather than waiting to be approached. A high AI score is a prompt for a conversation, not a finding of misconduct, and arriving at the conversation prepared changes how it goes. Ask what the score was, what the department's threshold is, and what evidence they will consider — and take notes.
Be able to talk about your own argument without the document in front of you. Why this research question, why this method and not the obvious alternative, where a particular source came from, what you changed after your supervisor's second round of comments. Someone who wrote the work can answer these fluently and someone who did not cannot, and in practice this matters more than any report.
Stay factual and unemotional. Describe your process, produce your evidence, and avoid attacking the tool — the department did not build it and an argument about detector reliability lands better as a documented limitation than as a complaint. If you used AI for anything legitimate, say so precisely and up front: an accurate account of grammar checking or search-term suggestion is far safer than a blanket denial that later turns out to need qualifying.
If your institution requires an AI-use declaration, complete it accurately rather than defensively. Declared, ordinary tool use is normally fine; an inaccurate declaration is the thing that causes real trouble.
If you have been flagged on work you wrote, send us the document and the report. We will give you an honest read on whether it looks like a detector artefact, what in the prose is likely driving it, and what evidence to take to your department.