Why AI detectors flag human writing
Not a bug that a better version will fix. It follows directly from what these tools are, and some human writing genuinely has the properties they measure.
Every AI detector produces false positives. Not occasionally, and not because somebody shipped a defect. It follows from what the method is: a classifier estimating a probability from statistical properties of text. Some human writing has those properties.
Ours does it too. The useful question is not whether a detector is ever wrong about human writing, it is how often, and on whose writing.
Detectors do not recognise generated text
They measure how closely writing matches the statistical profile of generated text. That is a different claim, and the gap between the two is where all the trouble lives.
Three signals do most of the work:
Predictability. How expected each word is given the ones before it. Models pick likely continuations, so their output averages out more predictable than prose a person wrote.
Rhythm. Variation in sentence length, punctuation density, clause structure. Human writing varies much more.
Word and phrase distribution. Which terms show up, and how often.
Text that is predictable, evenly paced and formally worded scores as machine written. Nothing in that sentence requires a machine to have been involved.
The writing that sits closest to the line
Non-native English. Learned English tends toward more regular construction and a narrower vocabulary range, both pushing toward the predictable end. Research keeps finding elevated false positive rates here, and we go into it in why non-native English writing gets flagged.
Technical and scientific prose. Methods sections are formulaic by design. Field conventions constrain vocabulary and sentence shape. A well-written methods section looks statistically similar to generated text because both are regular, and one of them is regular on purpose.
Anything written to a template. Structured abstracts, grant sections, lab reports. The template supplies the regularity for free.
Heavily edited writing. Editing smooths prose. Shorter clauses, consistent terminology, digressions removed. Smoothing strips out exactly the variation detectors read as human, which means the more work a piece has had, the more machine-like it can look.
Short text. A few hundred words carries almost no signal, which is why sentence-level error rates run several times document-level ones for every detector we know of.
Why nobody can engineer this away
A detector is separating two distributions that overlap. Wherever you put the threshold, some human writing sits on the wrong side of it. Move the line to catch less human writing and you catch less generated writing too.
That trade is the entire design decision. Most tools settle it by optimising average accuracy, which quietly accepts more false positives in exchange for a better headline number. We settle it the other way: fix the false positive rate first, take whatever detection rate falls out. At our public setting that is 0.8% of pre-LLM academic writing flagged, 0.4% at a stricter one.
That is a preference about which error to make. Not a claim to have removed either.
What a score can and cannot carry
It describes how text reads. It carries nothing about how the text came to be.
Several very different situations land in the same place: a model generated it, you drafted it and a model edited it, a model drafted it and you edited it, or you wrote every word yourself in a formal register. The words hold no record of the process.
A score describes how writing reads, which is exactly what an institution's tool is scoring too. That is the argument for knowing what yours says before somebody else runs theirs.
Lowering your own odds
If your writing keeps getting flagged and you want fewer surprises:
- Vary sentence length on purpose. Strongest single signal, easiest to change, costs you
no meaning at all
- Cut leading connectives. "Moreover", "Furthermore", "Additionally" raise the score and
almost never carry weight
- Keep drafts. Version history is a record of process. Finished text never is
- Check before you submit, not after. Ours is free and unlimited and scores each
paragraph separately, so you can see which passages carry the signal rather than staring at one number