You got an AI score. Here is how to tell whether it means anything.

7 min read · Published October 2026 · The third part of our detector series. Part one | Part two

Why a single number is not an answer

Every detector returns a percentage, and percentages feel like measurements. They are closer to weather forecasts: derived from real signals, useful in aggregate, and wrong about specific days.

In our own testing the same detector gave a genuinely human-written memoir 100% on one passage and 0% on a work email written the same week. That is not a rounding error. It is a range wide enough to cover almost any claim you might want to make.

So the question is never "what did it score". It is "does this score survive the checks below".

Check 1: How long is the text?

Under 100 words, stop. The number is noise. We watched human-written passages score 100, 100, 78.3, 0 and 0 — the same person, the same week, five documents.

Between 100 and 200 words, treat the score as a hint. Above 300 words it starts behaving like a consistent measure of style, which is still not the same thing as authorship.

Check 2: Run it twice

If the two numbers differ by more than 20 points, the tool is not measuring a property of your text — it is sampling. Identical input, identical settings, three runs in our test: 46.5, 0 and 0.

A spread that wide means the honest reading of "your text scored 40%" is "your text scores somewhere between 0 and 60 on Tuesdays".

Check 3: Test with a control

This is the check almost nobody runs, and it is the one that settles the question. Take something you wrote years ago, before generative AI existed. Run it through the same detector.

If that scores 50% or higher, the tool has a floor you cannot get under by writing better. In our sample, four of six human-written passages scored above 50%, and the average human score was 58 out of 100. Nothing you do to your own draft will read as "safer" than that baseline.

Check 4: Ask who is selling what

If the detector and the humanizer sit in the same dashboard, the rewriter will always appear to work — the vendor is grading its own homework. That is not necessarily dishonest; it is simply not evidence.

The same applies to review sites, including this one. We run a humanizer, which is why we publish our failures: the academic mode that barely moves, the student essay that went from 100% to 96.6%. Judge us by the number we could not fix.

Check 5: What decision is this score being used for?

This is the only question that really matters, and it is about power, not statistics.

If you are the one being accused

This happens constantly, and the people it happens to are usually not cheating — they are simply writing in a clean, even register, which is what the detector is trained to notice.

1. Ask how the score was produced. Which tool, which version, what length, how many runs. Vague answers are common and useful to you.
2. Run the control yourself. Take a document you wrote before 2022 and put it through the same tool. Bring the result with you.
3. Bring your drafts. Version history in Google Docs, Word, or Notion timestamps beats any detector in every process we have heard of.
4. Separate the two claims. "This reads like AI" and "you did not write this" are different statements. Concede the first if you have to; the second needs evidence the score cannot provide.

The short version

A detector score is a style opinion with a decimal point. Useful when you are editing. Dangerous when someone is deciding something about you. Check the length, run it twice, test a control you know the origin of — and then decide how much weight the number deserves, which is usually less than it is being given.

The data behind this series: nine samples scored against their known origin, and correlations across nineteen texts. Raw files and scripts are linked from each.

Frequently asked