When the detector mistakes a human for AI
It happens, and it happened to us. An article written by a person got a high "AI" score, while a machine-written text looked more human to the detector.
If it happens to you, the detector is not broken. It has shown its limits.
What happened
The owner of the service checked their own article, written without any AI, and got a high "AI" percentage. Then they checked a text the AI wrote from a free-form request and got "more human". The first reaction is to assume the tool is broken. The second is to see that the tool had just shown its limits.
It is not an exception
In our measurements across Russian, Ukrainian, Portuguese and English, the share of AI texts flagged above the threshold depended heavily on the service: from roughly 20% to 90%. At the same time, among neatly edited human news texts in Russian and Ukrainian about a third got high scores. The same threshold catches some and misses others.
Why it happens
The detector sees style: an even rhythm, typical phrases, no rough edges. A person who edits an article for a long time makes it even. An AI told to write simply and informally adds "liveliness". Style does not care who produced it.
Two more things work against a single text. First, length: below about 150 words our result is flagged as unreliable. Second, genre: promotional and social-media writing makes the signal weak, and the service says so with an abstention note instead of a confident number.
What it means for people who write
People start to fear their own tidiness. Some simplify their writing, or even make it sloppy on purpose, just to avoid a label. Students, freelancers and anyone writing in a second language are suspected most often. Writing well should never be grounds for an accusation.
If you were flagged
- Do not argue about the percentage. Offer the process instead: drafts, version history, notes, the sources you read.
- Offer to explain any passage in your own words. Authors can usually do it quickly.
- Check your own text and open the evidence list. It shows the exact lines that look uniform or contain stock phrases, so you know what the other side saw.
- Keep your drafts from now on. It costs nothing and settles most disputes.
If you are the one who flagged
A score is a reason to reread, not a verdict. Ask for drafts, talk to the author, and decide only when several independent things agree. The fair checklist for teachers and editors gives the order of steps.
What to do next
If you are unsure about your own text, check it and read the highlighted lines. To understand the reasons behind the errors, read why modern AI detectors get it wrong. If you want to bring the text closer to your own voice, see how to make your text sound more alive. Our service tries to show uncertainty and concrete places to fix rather than hand down a verdict.