The confusion is understandable
Both tools take text and return different text with a similar meaning. Under the hood the operation is close enough that several products sell the same engine under both names, and a reader comparing feature lists will find very little separating them.
The distinction that actually matters never appears in a feature list, because it is about where your input came from. That is a fact about your process, not about the product, and it decides everything downstream.
Paraphrasing puts you in a citation situation
If the passage you pasted came from a source, you are restating that source — and restating a source without attributing it is plagiarism, at any level of rewording. The similarity score falls because string matching is what it measures, which is a fact about the detection method rather than about the borrowing.
Done properly this is ordinary academic work. You read a source, you express its point in your own words because that is clearer than a block quotation, and you cite it. The tool can genuinely help with the middle step. What it cannot do is supply the citation, and it has no way to know whether you added one.
The full version of this is on whether paraphrasing removes plagiarism, including the case people find most surprising: following a source's structure and reasoning in your own words is the same offence, and it is harder to defend because it looks deliberate.
Rewriting your own draft raises no such question
If what you pasted is yours, there is nobody to credit. The ideas, the evidence and the argument were already yours and what changed is phrasing, which is why most institutional policies treat this closer to a grammar checker than to anything else.
It is not universally permitted — some courses prohibit any AI tool in any capacity — but the question you are answering is about your course's rules rather than about attribution. Those are covered on whether using one is cheating.
What detectors make of each
Not much that helps you. Both operations change the statistical shape of the prose, which is the thing a detector measures, and no classifier can see why a passage was rewritten. The research on this is usually cited at half length — the paraphrase attack paper did show detectors failing, and it also proposed a retrieval-based defence that worked.
Several vendors now advertise checks aimed at rewritten text. They do not, and could not, distinguish a student improving their own paragraph from one disguising a borrowed one.
The question to ask yourself
Where did this text come from? If a source, cite it, and use the tool to help you say the point clearly rather than to make the wording unrecognisable. If your own draft, you are editing, and the only remaining question is what your institution permits.
That question is answerable in ten seconds and it is more reliable than any comparison of what two products call themselves.