Why the same text scores differently the second time
Detectors that ask a language model are not fully deterministic — the same text can come out slightly differently. Two more things change between checks on their own: the history your text is compared against, and any small edit you made. A few points either way is normal; a big jump means the text or the material changed, not that the tool is broken.
What varies on its own
Some detectors ask a language model, and it does not answer identically every time. A small spread between runs is a property, not a fault.
Detectors that compute properties of the text directly barely move. That is another reason to read all the numbers side by side.
What changes between checks
Style deviation compares the text with your earlier texts. As more of them accumulate, the baseline it measures against shifts.
If you edited the text in between, even a small change in sentence structure moves the numbers. Rewriting one paragraph shows.
How big a difference to take seriously
A few points up or down changes no conclusion. On borderline texts the swing grows, because very little is needed to tip them.
When the difference is large, compare the flagged sentences. Either they point elsewhere, and then the text changed, or at the same places, and then it is only measurement spread.
FAQ
Should I check a text several times?
On borderline results it makes sense as a probe. Averaging numbers from several runs does not help.
Does the variation mean nothing can be trusted?
No. It means a single number cannot be trusted to the decimal. The direction and the flagged sentences are steadier than the exact value.
Why does a check show a change from the last one?
DetekceGPT shows the difference from the previous check so you can see whether the text moved. On an unchanged text, expect a small number.
Try it on your own text
Paste a text or upload a file and look at the score and at the specific sentences that came out suspicious.
Run a detection