Why AI detectors flag human writing
The false positive problem is structural, not a tuning issue. Who it hits, why, and what to do if it happens to you.
Detection · 2 min read
If a detector has flagged something you wrote yourself, you are not an unusual case. The false positive rate on these tools is high enough that the outcome is expected rather than surprising.
Why it is structural
Detectors measure how predictable text is, then infer authorship from that. But predictability and authorship are different things that merely correlate. Any human whose writing is predictable for legitimate reasons gets caught.
That includes:
Non-native English speakers. Writing in a second language means drawing on a smaller active vocabulary and more regular grammatical patterns. That is genuinely lower perplexity. Detectors read it as machine-written. This is the best-documented failure mode and it falls hardest on the people least able to contest it.
Technical and scientific writing. A methods section is supposed to be formulaic. So is a legal clause, a clinical note, a standards document. The register demands predictability.
Anyone well edited. Copy editing systematically removes the surprising choices that raise perplexity. A polished draft looks more machine-like than a messy one. The better your editor, the worse your score.
Short passages. Under a few hundred words there is not enough signal for any statistical measure. Scores on short text are close to noise.
Why "it said 95%" means less than it sounds
That number is not a probability that a machine wrote your text. It is a score describing how unusual your text looked to that tool's reference model, mapped onto a 0-100 scale by a threshold somebody chose.
Run the same passage through three detectors and you will often get three incompatible answers. That is not one of them malfunctioning — it is what happens when tools measure a proxy rather than the thing itself.
If you have been flagged
Practical and in rough order of usefulness:
- Ask what tool and what threshold. A specific number from a specific tool is contestable. "The detector said so" is not.
- Produce process evidence. Version history, drafts, commit history, a document's revision timeline. This is far stronger than any argument about the text itself, and it is why keeping drafts matters.
- Ask for the same tool to be run on known-human writing of yours. If your older, definitely-human work also scores high, that establishes the tool's behaviour on your voice specifically.
- Cite the false positive research. The bias against non-native writers is documented in published work, not merely anecdotal.
What cleaning does and does not do
Removing hidden characters and normalising punctuation is exact and verifiable, and worth doing on any text you did not write yourself. It has nothing to do with detector scores.
Rewriting stylistic markers raises perplexity, which will tend to move a score. But nobody can promise a particular result from a particular tool, and any product that does is guessing on your behalf.
If your text is your own and it got flagged, the honest advice is that the tool is wrong and process evidence is your strongest response — not that you should rewrite your own voice to satisfy it.
Related: how AI detectors work