The instruction sent to the model with every paraphrase includes the sentence "Keep names, numbers and quotations exactly as they are." I put a passage through it in three tones to see how well that holds. The names survived all three. Most of the numbers survived. The quotation did not survive once. In the original, a researcher described a dropout rate as "well within what we planned for"; in the three rewrites she says it was "consistent with the team's initial estimates", that it was "manageable", and that it was "considered okay". The quotation marks are gone in all three, which is the only mercy in it, because the words inside them are no longer hers.
TL;DR: A paraphraser is reliable at the thing that is easy to check and unreliable at the thing that matters. Measured on one 87-word passage in three tones: paragraph count held three out of three, length stayed within 10% of the original, and 44 to 55% of the original's words were replaced, so the rewording is real. Names, participant counts, percentages and most figures survived. A quoted sentence was reworded in all three, one currency figure vanished in two of them, and a date became American. Rewrite your own prose with it; do not rewrite anyone else's words. The paraphraser runs on the site's own server.
What does a paraphraser do that a thesaurus does not?
Change the structure, not just the vocabulary.
Swapping words for synonyms produces text that is detectably the original with a costume on, and often nonsense, because synonyms are rarely interchangeable in context. A paraphrase worth the name re-expresses the sentence: it changes clause order, splits or joins sentences, moves from passive to active, and chooses different words as a consequence rather than as the operation.
That is visible in the measurement. Across three tones, between 44% and 55% of the distinct words in the original did not appear in the rewrite, and the sentence boundaries moved. The output is not the input with substitutions.
What has not changed is the meaning, and that is the thing you have to check yourself.
What held up across every tone?
The structural promises, which are also the easy ones.
| Measure | Original | Professional | Casual | Simplified |
|---|---|---|---|---|
| Words | 87 | 96 | 91 | 88 |
| Length ratio | 1.00 | 1.10 | 1.05 | 1.01 |
| Paragraphs | 3 | 3 | 3 | 3 |
| Words replaced | - | 44% | 55% | 53% |
Three paragraphs in, three paragraphs out, every time, and never more than 10% longer. The prompt asks for both and gets both.
This matters more than it sounds. A paraphraser that silently merges your paragraphs or doubles your word count makes you re-edit the structure you had already settled, which is most of the work you were trying to avoid.
What survived and what did not?
Here is every checkable fact in the passage, against all three rewrites.
| Fact | Professional | Casual | Simplified |
|---|---|---|---|
| Name: Dr. Amara Osei | kept | kept | kept |
| Date: 14 March 2026 | kept | kept | kept |
| 1,248 participants | kept | kept | kept |
| Nine sites | kept | kept | kept |
| Dropout 7.3% | kept | kept | kept |
| Cost £1,249.50 | kept | lost | lost |
| Setup charge £48,215 | kept | kept | kept |
| Deadline 30 June | kept | kept | kept |
| Phase ends 2027 | kept | kept | lost |
| Quotation, verbatim | lost | lost | lost |
| Quotation marks at all | lost | lost | lost |
Professional lost nothing but the quotation. Casual dropped the cost per participant. Simplified dropped the cost and the end date.
The pattern is worth naming: the looser the tone, the more it drops. Casual and Simplified are being asked to sound relaxed, and a relaxed sentence carries fewer figures. The model resolves the conflict between "keep every number" and "sound like a friend" by quietly favouring the tone.
So the practical rule is that the tone setting is also a fidelity setting. Professional and Academic keep more than Casual and Simplified do, and if the passage is dense with figures, that is the reason to pick one.
Why is the quotation the serious failure?
Because a reworded quotation is a false statement about what somebody said.
In the original, the researcher described the dropout rate as "well within what we planned for". Here is what the three rewrites have her saying:
Professional ... was consistent with the team's initial estimates.
Casual ... and she said it was manageable.
Simplified ... which was considered okay.
Each is a plausible gloss. None is what she said. The Casual version is the worst, because "she said" attributes the new wording to her directly.
Dropping the quotation marks is the one thing that stops this being outright fabrication, and it is not a safeguard, it is an accident of the rewriting. A journalist, a student or a researcher pasting a passage containing a source's words and taking the output at face value has published a misquotation.
This is not a flaw in this particular tool. Rewriting is what a paraphraser does, and a quotation is the one span of text that must not be rewritten; the two requirements are in direct opposition inside a single instruction. The tool now carries a standing caution under the output saying so, written by the page rather than asked of the model, because asking the model for a caution gets you one about half the time.
Take quotations out before you paraphrase, and put them back afterwards. That is the whole fix and it takes ten seconds.
Does it change meaning as well as words?
Sometimes, and Simplified is where it showed.
The original says "The dropout rate was 7.3%". The Simplified rewrite says "About 7.3% of the participants didn't want to join". Those are different claims: dropping out of a trial after enrolling is not declining to enrol, and the second describes a refusal that never happened.
In the same rewrite, "The committee asked whether the cost per participant could be reduced" became "The group decided if the cost for each person would be lower". Asked became decided, which converts an open question into a resolution.
Both are the characteristic failure of simplification: the simpler sentence is easier to read and says something slightly different. It is the mode to trust least on anything precise, which is awkward, because precise technical text is exactly what people reach for Simplified to handle.
Which tone should you pick?
By what you are protecting, not by how you want it to sound.
| Tone | Keeps detail | Good for |
|---|---|---|
| Professional | Best in test | Anything with figures; business writing |
| Academic | Expected high | Formal prose, where register matters more than warmth |
| Formal | Expected high | Correspondence |
| Creative | Lower | Marketing copy where facts are few |
| Casual | Lost one figure | Internal notes, informal messages |
| Simplified | Lost two, drifted | Explaining a concept, not reporting a result |
Only Professional, Casual and Simplified were measured here; the other three are placed by what they ask the model to do. The ordering follows from the conflict described above: the further a tone pulls from the original register, the more detail gets traded for voice.
Why does the same text come back differently each time?
Because the setting that makes rewriting possible also makes it variable, and this tool is deliberately in the middle of that range.
Every AI tool here carries a temperature. The naming tools run at 1.0 because a repeatable list of names is useless. A grammar checker would want something near zero. The paraphraser sits at 0.5, which is below the default of 0.7: enough variation to produce genuinely different phrasing, little enough that it does not wander.
The practical consequence is that if a rewrite loses a figure, running it again will often keep it. If a sentence comes back wrong, a second attempt is cheap and frequently better. That is a reasonable workflow and it is also an argument against trusting any single output.
Is this the same as an AI detector arms race?
No, and treating it as one is how people get into trouble.
Paraphrasing text to evade an AI detector does not work reliably, and the premise is wrong in both directions: detectors produce false positives on human writing and false negatives on machine writing. One widely cited study found that detectors consistently misclassify non-native English writing as machine-generated, which is the clearest demonstration that the signal they read is not the one they claim to read. Running a passage through a paraphraser to change its statistical fingerprint is a gamble on a system nobody can audit, with your credibility as the stake.
The legitimate uses are the ordinary ones. Rewriting your own sentence that has come out clumsy. Shifting your own draft into a different register for a different audience. Producing a plain-language version of something you wrote in technical language. In every one of those, the text is yours and the paraphraser is an editing tool.
Rewriting someone else's text so it reads as yours is plagiarism whether a machine did the rewriting or not. The change of words is not what makes it not plagiarism.
How do you paraphrase properly for an essay or a report?
By understanding the passage and writing it again, which is the part a tool cannot do for you, and then citing it, which is the part people forget because the words changed.
The method every writing centre teaches is the same one: read the passage, put it away, write what it said from memory in your own words, then check your version against the original for accuracy. Purdue's guidance on quoting, paraphrasing and summarising sets out the standard version. Putting the source away is the load-bearing step, because a paraphrase written with the original in front of you tends to follow its sentence shapes.
A machine paraphrase inverts that. It has the original in front of it and nothing else, so it produces exactly the sentence-shaped echo the method is designed to avoid. The output changes the words and keeps the architecture.
Which leaves two honest uses in academic work. Rewriting your own clumsy sentence, which is editing. And producing a rough plain-language version of a dense source as a step towards understanding it, before you close the tab and write your own. Neither of those is the thing most people are searching for, and the one most people are searching for still needs a citation.
Is there a way to check the rewrite kept everything?
Yes, and it takes less time than reading the two versions side by side.
Copy the original and the rewrite into any text comparison tool and look only at what was deleted. You are not interested in the new wording; you are interested in whether a figure, a name, a date or a quoted phrase appears on one side and not the other. That is the check that caught every failure in this guide.
A rough version by hand: search the rewrite for each number in the original, one at a time. An 87-word passage has perhaps eight of them and the search takes a minute. That minute is how the missing £1,249.50 was found in two of three rewrites.
What no tool will flag for you is the verb drift, because "asked" and "decided" are both present, both plausible and both in the right place. That one needs a human reading for meaning rather than for difference, which is the argument for keeping a paraphrase short enough that you are willing to read it properly.
How long does a paraphrase take?
Under a minute for a short passage, and the time tracks length rather than difficulty.
The three runs took 60, 55 and 51 seconds for an 87-word passage, producing 622, 502 and 491 bytes. A small model on a CPU-only server writes at a roughly constant rate, so a page of text takes several minutes rather than several seconds.
Those timings also sit against a shared model: several AI tools on the site queue against one instance, so a busy moment adds waiting before any writing starts.
For a paragraph or two this is unremarkable. For a long document, paraphrase section by section, which is better practice anyway because it keeps each check small.
What should you check in the output?
Four things, in the order they go wrong.
- Anything in quotation marks in your original. Expect it to be gone. Restore it by hand.
- Every number and date. Professional kept all of them in test; the looser tones did not, and a date can come back in a different convention.
- Verbs that carry a claim. Asked, decided, suggested, confirmed. These shift in simplification and the shift is invisible unless you compare directly.
- Paragraph breaks. They held in all three runs, but they are load-bearing for anything you are pasting into a layout.
Read the rewrite beside the original rather than on its own. A paraphrase reads fluently whether or not it is faithful, which is precisely why it needs the comparison.
Why does it run on the site's own server?
Because the text people paraphrase is often not public and often not theirs to hand around.
Covering letters, draft contracts, a complaint before it is sent, a paragraph from a report under embargo, a medical summary being put into plain language for a relative. The usual alternative pastes it into a service that logs prompts, may train on them, and belongs to a company you have no relationship with.
Here it goes to a small model on the site's own hardware, and it is not stored unless you choose to send feedback. The cost is in the numbers above: a minute for a paragraph, and output that needs reading rather than trusting.
