Does GPTZero detect paraphrased AI text?
Partly
Often, but far less reliably than it detects raw model output.
How it works
Paraphrasing raises perplexity by swapping in less predictable synonyms, but usually leaves burstiness flat, because clause structure survives the rewrite. Detectors that weight both signals still catch a lot of it.
The part people miss
'Often' is not 'always' and not 'never'. Anyone quoting you a percentage here is guessing.
Flagged for something you wrote yourself?
There is a right order to handle it in, and editing the document first destroys your best evidence.