How AI detection works, and where it goes wrong
A plain-language guide to reading a detector's result sensibly.
What a detector actually measures
A detector cannot see who wrote a text. It can only look at the text itself and ask whether it resembles the kind of writing language models tend to produce. That is a question about patterns, not about authorship, and it is the reason every result is an estimate.
Signal one: rhythm
People write unevenly. A long sentence that wanders is followed by a short one. Then a fragment. Language models are far more regular: their sentences tend to be similar in length and built the same way. Measuring how much sentence length varies across a text, sometimes called burstiness, is one of the oldest and most useful signals.
Signal two: predictable wording
Models choose likely words. The result reads smoothly, but it also leans on the same safe phrases again and again: stock transitions at the start of sentences, tidy lists of three, and a closing sentence that sums up what was just said. A text that is full of these habits looks more machine-made, in any language.
Signal three: a second reader
Statistics miss things a reader notices, such as a lack of concrete detail, opinions that never commit to anything, or a tone that stays perfectly even from start to finish. For that reason TxtChecker also asks two separate AI models to read the text and judge it, and it combines their views with the statistical analysis. When the three agree, confidence is higher. When they disagree, the result says so.
Why detectors make mistakes
- Short texts. A few sentences do not contain enough pattern to measure. Results on short texts are unreliable.
- Formal writing. Legal, academic and technical texts are regular and polished by design, so careful human writing can look machine-made.
- Second-language writers. People writing in a language they learned later often use simpler, more predictable phrasing, and are flagged more often than they should be.
- Edited AI text. A text that was generated and then rewritten by a person sits between the two, and detectors disagree about it.
- Language. Most detection research is on English. Results in other languages are less dependable.
How to read a result
Treat the percentage as a likelihood, not a verdict. A high number means the text shares many patterns with machine writing. It does not mean a person did not write it. Look at the confidence level, read the highlighted sentences yourself, and ask whether the explanation matches what you see.
If the result matters, for example in a classroom, use it to start a conversation. Ask about the writing process, earlier drafts and sources. A detector's score should never be the only evidence against someone.
Writing that sounds like you
The habits that make text look machine-made are also habits of dull writing. Vary your sentences. Prefer the specific word to the general one. Cut transitions you do not need. Say what you actually think, and include the detail only you would know. These are good rules whether or not anyone is checking.