"Every AI Detector Gives Me a Different Result" — Which One Do I Trust?

The question

I put the same text into three tools and got 12%, 58% and 94%. Can they really differ that much? Which one am I supposed to believe?

Short answer

None of the three is the kind of number you "believe" outright. And a spread that wide isn't an exception — it's ordinary.

Why they differ this much

1. They weigh different things

Each tool decides for itself which aspects of style matter and how much. Looking at the same text with different emphases produces different results. It's less like measuring temperature and more like each tool grading "how smooth is this?" on its own scale.

2. They learned from different text

A tool built mainly on one kind of writing behaves unpredictably on another. This is the documented reason detectors flag non-native English writers so often: plainer vocabulary and shorter sentences look, statistically, like the thing the tool was taught to catch.

3. They're sensitive to length

The shorter the text, the bigger the spread. With three or four sentences there isn't enough basis for a statistic. That's why pasting the first half and the second half of the same document separately gives you different scores.

4. The result isn't a probability

The most misread point. "94%" does not mean "94% chance AI wrote this". It's closer to the strength of a signal, and each tool decides which scale to put that strength on. They were never comparable numbers to begin with.

So what should you look at

Don't try to average the numbers. Look instead at the spans that several tools flagged in common.

If all three landed on the same paragraph, that paragraph genuinely does read flat. Scores differ, but judgments about where the problem sits often overlap.

Then open that paragraph and ask:

Is there any judgment or concrete fact of mine in here?

That is far more useful information than a score.

Don't use it as your defense

People who've been questioned sometimes bring back the result from whichever tool gave the lowest number. It almost never works. The other side knows the results are all over the place, and the fact that you ran several tools can itself leave a bad impression.

You defend with process, not scores. Drafts, version history, source material. The detail is in accused of using AI.

If you need a way to choose a tool

Choosing on accuracy is hard, because there's no way to verify it. These are more practical:

• Does it show results sentence by sentence (a tool that gives only a total is of little use)
• Was it built for the language you actually write in
• Does it avoid stating results as fact (be careful with a tool that declares "written by AI")
• Is it clear about what happens to the text you paste in

The AI detector here gives sentence-level analysis and reports its result as a reference AI probability. There's more on reading results in the FAQ.

Finally

No tool can identify an author. Rather than worrying about the gap between scores, the habit of opening the flagged spans is what actually improves the writing.

Why making the score your goal costs you is covered in chasing a lower AI score will wreck your writing.