Half yours, half AI: what a detector shows on mixed text
On mixed text a detector shows something in between, and the overall number loses its meaning: one generated chapter is diluted across a fifty-page thesis into a score that looks innocent. What works is reading in parts — chapter by chapter, and then the sentences the tool flagged. Even then the detector will not tell you exactly where the author stops and the model starts.
The average hides what matters
The score is computed from the whole text you paste in. If ninety percent of the work is yours and one chapter is generated, the result leans towards the larger part and everything looks fine. That is how most missed cases in long documents happen.
The opposite is just as common: someone writes the whole thing themselves but takes a methods passage from a standard, and that pulls the score up. In both cases the culprit is a single number for the whole document, not the detector.
Editing by hand changes the signal
If you go through generated text and swap words, the score usually drops, but the sentence structure and rhythm stay. If you rewrite the structure — merge sentences, reorder them, add your own example — it drops further, because that is exactly what detectors track.
This is why results on mixed text diverge between tools so often. A detector that looks at phrases sees a clean text, while one that measures sentence rhythm still hears a metronome.
How to read such a text
Run the document in parts. A chapter gives a result you can work with, while the whole thesis at once only averages everything together. The colour map of the text then shows where flagged sentences cluster.
The second layer is a comparison with how the author writes otherwise. A passage that departs from their own earlier texts is worth a question — even when the overall score is low.
FAQ
Will a detector catch a single generated sentence?
In a short text perhaps, in a long one no. One sentence is diluted by the whole — running just that paragraph on its own helps.
Is mixed text automatically cheating?
No. It depends on the rules for that assignment. Usually it is enough to declare that you used AI and say what for.
How long should one part be?
The longer the passage, the more stable the result, but the greater the dilution. A chapter or a section is a reasonable compromise.
Try it on your own text
Paste a text or upload a file and look at the score and at the specific sentences that came out suspicious.
Run a detection