Perplexity is a search product, so its output is different in kind from a chat assistant's: it is a summary of sources, with the claims bound to references. Any Perplexity humanizer has to deal with that, and almost none of the pages on this topic mention it.
The risk is specific and it is serious. A rewrite changes wording to shift statistical properties. It does not know that sentence three is the one your reference supports. Change the claim and the citation is now attached to something the source never said, which is a worse problem than a detector score.
The reason this page is mostly about citations rather than about scores.
Perplexity gives you a claim and a source for it. The pairing is the value. A rewrite operates on the claim's wording with no model of what the source actually supports, so three things can happen and two of them are bad.
Strengthening. "Some studies suggest X may contribute to Y" becomes "studies show X causes Y". The reference is still attached. It now supports a claim nobody made. This is the most common and the most damaging, because the rewrite reads better and more confidently, so you are less likely to question it.
Softening. The reverse. Less dangerous, still wrong, and it quietly weakens an argument you had evidence for.
Invention. A detail appears that was in neither your text nor the source. We saw exactly this while testing for these pages: a pass returned the sentence "Roughly a year into it" when nothing in the input mentioned any timescale. On cited work, an invented specific sitting next to a reference marker is close to the worst output this category can produce.
Work a sentence or a short paragraph at a time, and after each pass read the highlighted changes against your source, not against the previous draft. The question is never "does this read better", it is "does the source still say this". That is slower than pasting the whole answer, and it is the only version of this workflow that is safe on referenced writing.
A practical detail that catches people on the first attempt.
The humanizer takes plain text and returns plain text. Inline markers that came through as characters, the bracketed numbers, usually survive as characters, because they look like part of the sentence. What does not survive is anything that was a link rather than a character: pasting into the box strips the hyperlink, so a marker can come back pointing at nothing.
The workable order is to rewrite first and attach references afterwards, from Perplexity's source list rather than from the rewritten text. If you re-link from the rewrite, you are trusting that the marker is still next to the claim it belonged to, which is the assumption this whole page is arguing against.
Also worth knowing: markers can confuse sentence boundaries, so a bracketed number mid-sentence sometimes gets treated as a break. For passages dense with references, strip the markers, rewrite, then reattach.
A property of the format rather than a flaw in the model.
Summary prose has a characteristic shape: even sentence lengths, each one carrying roughly equal weight, few asides, heavy use of "additionally" and "furthermore" to stack points. That is a reasonable way to summarise, and it is also close to a textbook description of what AI detectors read as machine-generated.
So Perplexity output often scores high for a reason that has little to do with it being generated: it is written like a summary, and summaries look like this. Human-written literature reviews and abstracts get flagged for the same reason, which is one of the better-documented sources of false positives. Who gets flagged unfairly is relevant if you are writing this way yourself.
A rewrite addresses it by varying the sentence rhythm and cutting the stacking connectives. Light does that and leaves your vocabulary alone, which is the right choice on referenced material. Balanced edits phrasing more broadly; Maximum replaces most of the wording, is a paid mode, and on cited work gives a rewrite the most opportunities to move a claim. For anything with references attached, lighter is safer.
It will not check your sources, tell you whether Perplexity summarised them correctly, or tell you what Turnitin will say about your document. Nobody can do that last one from outside Turnitin. And if you are already accused, do not rewrite: your draft history is the defence.
Yes. You can rewrite Perplexity output here with no account: 3 humanizations a day at up to 260 words, with an AI score before and after from our own detector and a highlight of which words changed. On citation-bearing text that highlight is the part that matters, because it is how you check whether a claim moved away from the source that supports it.
It can break the thing that matters more than the markers: the link between a claim and what its source actually says. A rewrite changes wording without any model of what your reference supports, so it can strengthen a hedged finding into a definite one while the citation stays attached. Separately, the markers themselves survive as characters but any hyperlink is stripped when you paste. Rewrite first, then attach references from Perplexity source list rather than from the rewritten text.
Largely because it is written like a summary. Summary prose has even sentence lengths, roughly equal weight per sentence, few asides and heavy stacking connectives like additionally and furthermore. That shape is close to what detectors read as machine-generated, which is also why human-written abstracts and literature reviews get false-positived. The format is doing much of the work, not just the fact that a model produced it.
A sentence or a short paragraph at a time, on the lightest setting, reading the highlighted changes against your source rather than against your previous draft. The question to ask is whether the source still says this, not whether the prose reads better. If references are dense, strip the markers first, rewrite, then reattach them from the source list. Pasting a whole cited answer and accepting the output is how a citation ends up supporting something nobody claimed.
Yes. In our own testing for these pages a pass returned the sentence "Roughly a year into it" when the input mentioned no timescale at all. On referenced writing an invented specific next to a reference marker is about the worst output this category produces, because it looks sourced and is not. This is the main argument for reading the word-level diff instead of the finished prose.
No, and that is a real gap rather than something to gloss over. We rewrite text and score it; we do not read your sources or verify the summary against them. If Perplexity misread a paper, a rewrite will faithfully preserve the misreading in better prose. Checking the summary against the original is work only you can do.
A casual register, and the tells it still leaves behind.
Read the guide →Block-structured output, and where detection actually matters.
See the guide →Referenced writing, hedging, and what must not move.
Read the guide →The check that catches a strengthened claim.
See the method →Three a day, no account. Read the highlighted changes against your source before you attach a reference.