How to See Exactly What ChatGPT Changed in Your Text

ChatGPT returns a rewrite, not a redline. Here is how to see exactly what ChatGPT changed in your text, word by word, with real change stats from two prompts.

To see exactly what ChatGPT changed in your text, paste your original on the left and ChatGPT's version on the right of a word-level diff tool, then read the highlights: red for what it removed, green for what it added. A normal chat reply comes back as a block of prose with no markup, so the diff is the only way to get a redline of the edit.

The interesting part is how much changes, and it is usually far more than the prompt implied.

Key Takeaways

• ChatGPT does not mark its edits in chat. Canvas has a Show changes view for edits it makes inside the canvas, but a pasted-in rewrite in a regular reply arrives clean, with no trace of what moved.
• "Fix my grammar" is a light touch. In our test, a 377-word draft came back 96.6 percent similar in word mode with 20 words changed.
• "Make it sound more professional" is a rewrite. Same draft, 40.9 percent similar, 429 words changed, 26 sentences rewritten, nothing left untouched.
• The similarity number tells you which side of the line you are on. Above 90 means editing. Below 60 means the model wrote it, whatever the prompt said.
• The diff shows what changed, not who wrote it. For that second question, either version goes to an AI detector, which needs a free account.

Why you cannot see the changes inside ChatGPT

When you paste a paragraph into ChatGPT and ask it to improve the writing, you get a paragraph back. Not a marked-up copy. Not a list of edits. A replacement. You can read the two side by side and squint, and most people do exactly that, and miss half of what moved.

There is one partial exception. According to OpenAI's help page for Canvas, the canvas editor has a Show changes icon that strikes removed text in red and highlights additions in green, plus previous and next version arrows. That helps if you did the whole edit inside a canvas. It does nothing for a rewrite that came back in a normal reply, for text you copied out and kept editing in Word, or for comparing what Claude or Grammarly did to the same draft. It also cannot be shared with anyone who does not have your ChatGPT account.

So for the common case, someone pasted a draft into chat and got a rewrite back, the redline has to come from a separate tool.

How to see exactly what ChatGPT changed, step by step

1. Keep your original. This is the step everyone skips. Before you send anything to ChatGPT, copy the draft somewhere safe. If you already lost it, check your document's version history; Google Docs and Word both keep one.

2. Get ChatGPT's version as plain text. Copy the reply. If it came back with markdown headings or bullets you did not ask for, that is itself a change worth seeing.

3. Paste both into a diff tool built for prose. Open the GPTOne diff checker, original on the left, ChatGPT's version on the right, and pick the Original vs AI rewrite mode. It runs in your browser, so neither text is uploaded, and there is no account to create.

4. Read the highlights by word, not by line. Red strikethrough is what the model cut. Green is what it added. A red word directly followed by a green word is a substitution, which is what most AI edits are.

5. Read the stats block. Words added, words removed, sentences rewritten, similarity percentage. Those four numbers tell you the size of the edit before you read a single highlight.

For the general technique, including why line-based tools like Diffchecker mislead on prose, see how to compare two texts online.

What we found: two prompts, one draft

We wrote a 377-word draft, a short opinion piece about a team dropping its daily standup, deliberately left in a handful of errors ("there turn", "your going", "thats what teams do"), and had an AI assistant rewrite it twice with the two prompts people send most often. Then we diffed each result against the original in word mode.

The grammar prompt did what it said. Twenty words changed across nine sentences: the typos, a few hyphens, contractions expanded. Two of the six paragraphs came through completely untouched. Read the diff and every highlight is a correction you would have accepted from a copy editor.

The professional prompt is a different animal. It changed 429 words in a 377-word draft, which is possible because it removed 215 and added 214, and it rewrote all 26 sentences. Not one paragraph survived intact. "Look at me like I have cancelled Christmas" became "the reaction is often one of surprise". "Worth more than any ritual" survived, barely. The argument is the same. The writing is not yours any more.

Nobody who sends that second prompt thinks they are asking for a ghostwriter. But that is what the diff shows they got.

Why the similarity number matters more than it looks

The similarity percentage in the stats block is the share of words the two versions have in common, computed as a Dice coefficient. It is not an AI score and it is not a plagiarism score. It is a measure of how much of your wording survived.

That makes it the most useful single number for the question underneath this whole post: did ChatGPT edit my text, or did it write a new one? The line institutions draw between assistive and generative use is exactly that line, and we went through where Turnitin sits on it in does Grammarly trigger AI detection. Above 90 percent, the model corrected you. Below 60, the model wrote it and you approved it.

One caveat on modes. Sentence mode counts a whole sentence as changed when any word in it changed, so the same grammar edit that scored 96.6 percent in word mode scored 61.1 percent in sentence mode. Both are true. Word mode answers "how much of my wording survived", sentence mode answers "how many sentences were left alone". Quote the mode with the number.

What to do with the diff

Take back the sentences you want. The point of seeing the edit is choosing. Accept the corrections, reject the rewrites that flattened your voice, and paste the result back. This takes five minutes with a diff and an hour without one.

Check how each version reads to a detector. A rewrite that scores 40 percent similar to your draft is mostly model text, and detectors will usually read it that way. The GPTOne AI detector scores text from ChatGPT, Claude, Gemini, DeepSeek and other models at 99.99 percent accuracy on text, up to 50,000 characters per scan. It needs a free account, which comes with free credits and no card. Scan your original and the rewrite and compare the two scores next to the similarity number.

Keep the diff as a record. If you are a student, a share link of the comparison is a timestamped account of what you wrote and what the tool did. Note that the link contains both texts, so send it only to someone who should read them.

If the rewrite was heavier than you wanted and you now want to know whether tools that reverse it actually work, our honest look at whether AI humanizers work is the next read. And for the paraphraser version of this question, where the tool is QuillBot rather than ChatGPT, see compare original and paraphrased text.

FAQ

Does ChatGPT show what it changed?

Not in a normal chat reply. Canvas has a Show changes view for edits made inside a canvas document. For anything else, paste the original and the rewrite into a word-level diff tool.

How do I compare my original text with ChatGPT's version?

Paste your draft on the left and ChatGPT's reply on the right of a diff checker set to word mode. Removed words show in red, added words in green, and the stats block gives a similarity percentage.

How much does ChatGPT change when I ask it to fix grammar?

In our test, about 5 percent of the words. A 377-word draft came back with 20 words changed and 96.6 percent similarity. Prompts about tone or professionalism changed over half the text.

Can a diff tell me if ChatGPT wrote my essay?

No. A diff shows what changed between two versions you supply. A low similarity between your draft and the rewrite tells you the model produced most of the final wording, but whether a text reads as AI-generated is a question for an AI detector.

Is my text uploaded when I compare it?

Not with the GPTOne diff checker. The comparison runs in your browser and nothing is sent to a server or stored.

Paste your draft and ChatGPT's version into the GPTOne diff checker and see exactly what changed. When the question turns to how each version scores, the AI detector is one click away with free credits on signup.