The score is confidence, not a portion
An overall score like 78% is the detector's confidence that the whole text is AI-written. It is answering “how sure am I?”, not “how much of this is AI?”. Those sound similar but they are not the same thing.
What people think 78% means: 78% of the words are AI.
What it actually means: the tool is 78% confident the text is AI-written.
How to read the range
High (roughly 80% and up)
The detector is fairly confident it sees AI. On long, untouched text this is often right. On short or edited text, still treat it with care.
Middle (around 40 to 70%)
This is the tool saying it is not sure. A middle score is the least useful result, and the easiest to misread as a firm answer.
Low (roughly 20% and under)
The detector sees the bumpiness of human writing. Reassuring, but a low score is not a certificate either.
Why the length of the text matters
A score on two sentences is close to a guess. There is not enough writing to measure the patterns the detector relies on. The same tool on a full page is far more stable. So always read a score together with how much text produced it. A confident number on a tiny sample is not confident at all.
Sentence-level scores are different
Some tools also highlight individual sentences as more or less AI-like. That is a separate view from the overall confidence score, and it is useful for seeing which parts look smooth. But it carries the same warning: a highlighted sentence is a hint to look closer, not a finding. A single flagged sentence in an otherwise human essay usually means nothing, since one smooth line is a coincidence, not a pattern.
Why two detectors give different scores
Paste the same text into two AI detectors and you will often get two different numbers. One says 30% AI, the other says 80%. People find this alarming, but it makes sense once you know what a score is.
Each detector was trained on a different pile of writing and tuned to weigh the signals differently. They are all guessing at the same hidden thing from the same surface clues, but they draw the line in different places. The disagreement is not a bug you can fix by finding the “real” detector. It is a reminder that every one of these numbers is an estimate. When two tools disagree sharply, the honest reading is “this text is genuinely uncertain,” not “one of them is right.”
What to actually do at each score
Reading a score is only useful if it changes what you do. Here is a sensible response to each range, whether you are checking your own writing or someone else's.
High score on long text
Worth a closer look, but still not proof. Read the writing yourself and, if it matters, ask about the process. Never act on the number alone.
Middle score
The least useful result. Treat it as "no clear answer" rather than a lean either way. Do not build any decision on a middle score.
Low score
Mildly reassuring, nothing more. A low score does not certify that a human wrote it, since edited AI text scores low too.
Any score on short text
Ignore the confidence. A couple of sentences cannot produce a reliable number in either direction.