Compare Original and Paraphrased Text: How Much Actually Changed?

Compare original and paraphrased text word by word and measure the rewrite: words swapped, sentences rebuilt, paragraphs moved, with real similarity numbers.

A paraphrase changes text in three ways: it swaps words, it rebuilds sentences, or it moves paragraphs. To compare original and paraphrased text, paste both into a word-level diff and read which of the three happened. We cover this separately in Can AI Detectors Detect Paraphrased Text. In our test, a synonym-level paraphrase kept 88.6 percent of the wording. A restructuring paraphrase kept 28.6 percent. Same draft, same meaning, very different documents.

That gap is the whole story, so here is how to measure it yourself.

Key Takeaways

• Three kinds of change, three different readings. Word swaps keep similarity high, sentence rewrites drop it fast, paragraph moves should not change it at all.
• Synonym-style paraphrasing changed 80 words in a 377-word draft. 88.6 percent similarity in word mode, but 22 of 26 sentences touched.
• Restructuring-style paraphrasing changed 504 words in the same draft. 28.6 percent similarity, every sentence rebuilt.
• QuillBot's percent change and a diff similarity score measure different things. One is the paraphraser's own report on its output. The other is an independent comparison you control.
• Similarity is not a plagiarism score and not an AI score. It is shared words over total words. Nothing more.

The three kinds of change

Before comparing anything, it helps to know what you are looking for, because a paraphraser produces three distinct types of edit and they mean different things.

Word swaps. "Stopped" becomes "quit". "Explain" becomes "clarify". The sentence structure survives, the rhythm survives, the argument is untouched. In a diff, these appear as a red word immediately followed by a green one, scattered through otherwise unmarked text.

Sentence rewrites. The clause order flips, an active sentence goes passive, two sentences merge. The meaning is the same but the sentence is new. In a diff, most of the sentence lights up, and the similarity number falls quickly because few consecutive words survive.

Paragraph moves. The paragraph is intact but somewhere else. A code-style line diff shows this as a block deleted and a block added, which looks like a rewrite. A prose diff labels it as moved and leaves the similarity alone, because no words changed.

Paraphrasing tools do the first two. Editors do all three. Knowing which you are looking at is what turns a wall of highlights into a judgement.

How to compare original and paraphrased text

1. Paste the original on the left and the paraphrase on the right of the GPTOne diff checker, a text compare tool built for prose rather than code. It runs in your browser, nothing is uploaded, and there is no account to make.

2. Use word mode. Line mode, the default on most diff tools, treats a re-wrapped paragraph as a rewrite. Word mode compares paragraph to paragraph first, then highlights only the words that moved inside each one.

3. Read the stats block before the highlights. Words added, words removed, sentences rewritten, paragraphs moved, similarity percentage. Those numbers classify the paraphrase before you read a word of it.

4. Then read the highlights for pattern. Scattered single-word substitutions mean a synonym pass. Whole sentences lit up mean restructuring. Both in the same document mean the tool was run in a heavy mode, or run twice.

The general method, including how the colours work, is in how to compare two texts online.

What we measured: two paraphrase styles, one draft

Paraphrasing tools sell their modes on a spectrum. QuillBot's Standard mode promises to keep meaning while changing wording, its Creative mode promises the heaviest rewrite, and according to QuillBot's own help page, its Statistics panel reports a percent change figure and the longest unchanged run of words for each result.

We wanted to see what the two ends of that spectrum look like in an independent diff. We took a 377-word draft, a short opinion piece about a team dropping its daily standup, and had an AI assistant produce two paraphrases: one synonym-level with the sentence structure intact, the way a standard paraphrase mode is designed to work, and one where every sentence was rebuilt, the way a creative mode is designed to work. We did not run QuillBot itself for this test, so these are illustrations of the two styles rather than measurements of one product. Then we diffed each against the original in word mode.

Three things stand out.

The synonym pass touched 22 of 26 sentences while keeping nearly 89 percent of the words. That is what a light paraphrase does: it changes something in almost every sentence and almost nothing in any of them. In a diff it looks like a light rash of red and green pairs. Read closely and the voice is recognisably the original author's.

The rebuild pass changed more words than the draft contained. 504 changes in 377 words is possible because it removed 276 and added 228. Sentence mode similarity was zero. Not one sentence survived. The argument came through, the examples came through, and the writing is a different person's.

And the sentence-mode column tells its own story. The light paraphrase scored 15.4 percent in sentence mode against 88.6 in word mode, because a sentence with one swapped word counts as changed. If a tool reports a single "percent changed" figure, ask which unit it counts. The number can honestly be 11 percent or 85 percent for the same edit.

How this differs from the paraphraser's own statistics

A percent change reported by the paraphraser is the tool grading its own work. It is useful, and it is also the number the tool has an incentive to make look impressive. A similarity score from a separate diff is a comparison you ran, on the two texts you chose, with a unit you picked. Keep both, trust the second.

The other difference is scope. The paraphraser's statistics exist only inside the paraphraser, for that session. A diff works on any pair of texts: the original against a QuillBot output, against a ChatGPT rewrite, against a Grammarly suggestion, against a student's resubmission. We covered the ChatGPT case in detail in how to see exactly what ChatGPT changed.

What the comparison cannot tell you

A similarity percentage between two texts you supplied is not a plagiarism verdict. Plagiarism detection needs a corpus to compare against, and a diff has only the two texts in front of it.

It is not an AI verdict either. A heavily paraphrased text has low similarity to its source. Whether it reads as machine-written is a separate question, and research says the answer is unstable. According to Krishna et al., 2023, paraphrasing sharply reduces the accuracy of AI text detectors, which is why paraphrasers get used to evade them and why detectors keep retraining on paraphrased text. Our own look at whether detectors catch humanized text goes through what survives a rewrite and what does not.

If you want that verdict, run the paraphrased version through the GPTOne AI detector. It scores text from ChatGPT, Claude, Gemini, DeepSeek and other models at 99.99 percent accuracy on text, up to 50,000 characters per scan, and needs a free account with free credits and no card. Run the original too. The pair of scores next to the similarity number is the full picture: how much changed, and how each version reads. For a head-to-head on the detection side, see GPTOne vs QuillBot's AI detector.

FAQ

How do I compare original and paraphrased text?

Paste the original on the left and the paraphrase on the right of a diff checker in word mode. Removed words show in red, added words in green. The stats block gives a similarity percentage and counts of words and sentences changed.

What does QuillBot's percent change mean?

It is QuillBot's own report of how much its output differs from your input, shown in the Statistics panel alongside the longest unchanged run of words. It measures the tool's output; an independent diff measures the same thing on any two texts.

What is a normal similarity percentage after paraphrasing?

In our test, a synonym-level paraphrase kept 88.6 percent of the words and a full restructuring kept 28.6 percent. Anything above roughly 85 means wording was swapped; anything below 50 means the sentences were rebuilt.

Does paraphrasing make text undetectable to AI detectors?

It lowers detection accuracy, according to published research, but not reliably to zero, and detectors are trained on paraphrased text in response. Rewrites usually land in an ambiguous middle band rather than clearly human.

Is the comparison private?

With the GPTOne diff checker, yes. The comparison runs in your browser, nothing is uploaded or stored, and a share link carries the texts in the URL itself, so only send it to someone who should read them.

Paste your original and the paraphrase into the GPTOne diff checker and read the stats block. When you want to know how the paraphrase scores, the AI detector is one click away with free credits on signup.