Ask the AI to Fix the Typos and It Rewrites the Note

Ask the AI to fix the typos in your note and you often get the typos fixed — plus edits you never asked for. Language models correct for fluency, not for the smallest change, and 2026 models still do. Name the edit types you allow, keep your original file, and word-diff the result before you accept it.

Here is what that looks like. A shorthand note, and a fluent "fix" of it, compared word by word with git:

$ git diff --no-index --word-diff note.orig.md note.ai.md
[-i-]{+I+} went to [-teh-]{+the+} store and [-buy-]{+bought+} milk. {+I+} need to call [-bob abt-]{+Bob about+} the [-thing-]{+matter.+}

The fixed line is written by hand to illustrate the pattern — it is not captured model output. The diff is real, run on git 2.43.0.¹ Read the brackets. teh to the is the typo you asked about. buy to bought is grammar, and probably welcome. {+I+} completes a fragment. abt became about. And [-thing-]{+matter.+} is a different word — one you never chose, now sitting in your note as if you had.

This is the unrequested rewrite. What an AI rewrite does to the words you search by covers the rewrite you asked for, the "improve this" button. Here you asked for less. Researchers in grammatical error correction have measured how much more you get since 2023, and a paper accepted to EMNLP 2026 Findings shows this year's models doing the same thing.

Two jobs share one word: correction

Researchers separate two tasks that everyday prompts blur. Minimal-edit correction fixes errors and leaves the rest of the sentence alone. Revision rewrites for fluency and allows any change that keeps the meaning. "Fix the typos" asks for the first; a language model left to itself tends to deliver the second.

Steven Coyne and colleagues at Tohoku University and RIKEN put the split plainly in a 2023 preprint on GPT-3.5 and GPT-4:

"The first is education for language learners, in which case easily comprehensible minimal edits are employed, with an emphasis on achieving grammaticality but otherwise leaving the sentence as-is. The other is a revision task in which a sequence is edited to sound fluent and natural, and any number or type of changes can be applied as long as the intended meaning, as interpreted by the editor, is preserved."²

The clause that matters is the last one — as interpreted by the editor. In a revision, the editor decides what you meant. In your note, the editor is now a model.

The trade between the two is measurable. Tao Fang's team at the University of Macau found that "the less a system adheres to minimal edits, the better the fluency performance of the generated sentences."³ Fluency is what a model is good at. Leaving things alone is not.

A second 2023 report, from Haoran Wu and colleagues at CUHK and Tencent AI Lab, described the same habit from the other side: "ChatGPT goes beyond one-by-one corrections. Specifically, it prefers to change the surface expression of certain phrases or sentence structure while maintaining grammatical correctness."⁴

The measured rate — and how small the study was

In one 2023 human evaluation, ChatGPT over-corrected 43.0%, 38.0% and 60.0% of short, medium and long sentences. Grammarly's free Correctness feature over-corrected 11.0%, 9.3% and 18.5%. The test covered 150 learner-essay sentences, three raters and gpt-3.5-turbo — a small study of an old model on someone else's writing.

The scope, in the authors' words: "We select the 50 longest, 50 medium-length, and 50 shortest sentences from the CoNLL14 test set and invite three postgraduate students with international study experience to conduct a manual evaluation."³ For Grammarly, "we only use its free and open Correctness feature."³ The paper is an arXiv preprint, version 1 only — no venue is listed on it.

"Over-correction" here has a precise meaning, and it is not "made a mistake." Fang's team defines it as a system's ability "to produce correct results beyond what is indicated by the reference correction."³ An over-corrected sentence can be perfectly good English. It is just more edited than a careful human corrector chose to make it.

The full panel, from the paper's Figure 2:³

Two details keep this honest. In the medium bucket, Grammarly's 9.3% ties GECToR exactly. And the authors note that "Grammarly also exhibits over-correction compared to GECToR and T5 large models."³ Every system edits beyond the reference sometimes — the language model just does it far more often.

The abstract's summary is fair to both sides: ChatGPT "has excellent error detection capabilities and can freely correct errors to make the corrected sentences very fluent, possibly due to its over-correction tendencies and not adhering to the principle of minimal edits."⁵

Not a 2023 artifact: the 2026 models do it too

A paper accepted to EMNLP 2026 Findings tested GPT-5.4, Claude Opus 4.6, Gemini 3.1-Pro and four other models on standard correction benchmarks. Given a plain prompt asking for grammar and spelling errors to be fixed, every model showed the same over-correction pattern — high recall, low precision. The behaviour moved forward with the models. It did not age out.

Kateryna Karpo (Ukrainian Catholic University and YouScan) and Artem Chernodub (Zendesk) posted the paper on 9 September 2026.⁶ Their result for the plain prompt: "The vanilla zero-shot prompt (A.1) shows a classic overcorrection pattern across all models, with high recall but consistently poor precision (often below 45.0)."⁷

High recall means the models found many of the real errors. Low precision means much of what they changed was not on the annotators' list — edits the human correctors never made. They name the extra edits directly: "rewriting well-formed spans with synonym swaps, stylistic polishing, or clause rephrasings that add no actual value."⁷

thing to matter is a synonym swap.

A peer-reviewed paper at EMNLP 2025 reached the same place from a different direction. Taehee Park, Heejin Do and Gary Geunbae Lee wrote that "Large Language Models (LLMs) often show the opposite tendency, making excessive overcorrection, leading to low precision."⁸

What this does not give you is a current rate. Nobody has published a 2026 over-correction percentage for your model on your notes, and this post will not invent one. The 2026 paper measured learner-English benchmarks, and its own limitations section warns that the public datasets "may have been encountered by commercial LLMs during pretraining."⁷ It is also version 1 of the paper; the published camera-ready may differ.

Why a note is the worst case

Correction benchmarks are built from essays by learners of English — full sentences, meant to be read by someone else. Notes are fragments, shorthand and lists — and fragments are what one 2023 study saw models expand. To a fluency editor, a fragment looks like an error, so the "fix" it makes can be new content rather than a corrected typo.

Coyne's team listed it among their findings: "The models occasionally over-edit, changing the meaning of a sentence during correction, or expanding fragments with new material."²

A note is mostly fragments, on purpose.

need to call bob abt the thing is a complete thought to the person who wrote it. "The thing" is a pointer only you can resolve. "The matter" is a model's guess that now reads like your own memory — and in six months you will not know which words were yours.

One caveat cuts in the model's favour. Coyne's experiments were run in a revision setting — in their words, "Due to our sentence revision setting, our experiments focused more on fluency edits than minimal edits."² Part of what they measured was what they asked for — and that supports the routine below rather than weakening it: the request sets the scope.

They also found the request is fragile. "Both the error correction quality and the reliability of the output format differ significantly based on simple changes to wording or punctuation."² Small changes in how you ask move the result. That is why the check cannot be the prompt.

The single habit: keep the original, diff what comes back

Never let the corrected text overwrite the note. Save what came back as a second file, run a word-level diff against your original, and accept changes one by one. The diff does not depend on which model you used or how well you phrased the request, which makes it the one step that survives every model update.

This is the same convention as keeping the original beside an AI summary: the model's output is a proposal, and your file is the source.

Apple, at least, treats the two jobs as separate actions. Its Writing Tools on macOS 27 list them as different buttons — "Click Rewrite. To choose a particular tone, click Friendly, Professional, or Concise." and, separately, "Click Proofread."⁹ For its ChatGPT extension, the same page notes you can "revert to the original version of your text."⁹ A revert button exists because the result is not always what you wanted.

A pasted note in a chat window has no such button. The second file is the revert.

The five-minute version

Seven steps, each under a minute once you have done it twice. Copy the note, decide which edits you allow, ask with those edit types named, save the reply as a separate file, word-diff the two, accept or reject each change, and reread once for the typos a careful request will miss.

1. Copy the note. cp note.md note.orig.md. Ten seconds. If your notes break lines at sentences or clauses, the diff stays readable — wrap width decides how a diff reads.
2. Decide what you allow. Spelling, punctuation, capitalisation, agreement — pick the list before you paste. Decide, too, where the text is going: your spell-checker is already a data-egress decision, and a chat model is the same decision, larger. Twenty seconds.
3. Ask with the edit types named. Use the prompt below, trimmed to your list. Twenty seconds.
4. Save the reply as its own file. note.ai.md, never over the original. Ten seconds.
5. Word-diff the two files. git diff --no-index --word-diff note.orig.md note.ai.md works on any two files, inside a repository or not. The mechanics — modes, colour output, why it splits on whitespace — are in a word-level diff of two text files. Ten seconds.
6. Accept or reject each bracket. Anything outside your list gets rejected by default — a synonym, a completed fragment, a reordered clause. One to two minutes.
7. Reread once for misses. A tighter request fixes fewer real errors; the next section has the numbers. One minute.

The whole routine is two files and one command.

The git manual describes the form in step 5 as the way "to compare the given two paths on the filesystem."¹⁰ By default, "words are delimited by whitespace,"¹⁰ so abt to about shows as one change, not a line. If the diff output is hard to read, the algorithm matters too — Myers, Patience and Histogram split text differently.

Here is a prompt for step 3. It is ours, built from the design the 2026 paper tested — a minimal-edit rule plus a named list of error types — but this exact wording has not been benchmarked:

Fix only these error types in the note below:
- spelling mistakes
- punctuation
- capital letters at the start of sentences and in names
- subject–verb agreement and verb tense

Do not rewrite, rephrase, reorder or complete sentences.
Leave fragments, shorthand and abbreviations exactly as they are.
Do not add or remove words except to fix an error from the list.
If a line has no error from the list, return it unchanged.
Return the whole note, in the same line order.

On the receipt note, a model following this prompt to the letter should return I went to the store and bought milk. Need to call Bob abt the thing — one typo, one tense, three capitals. Everything else stays yours, including abt and the thing.

What naming the edits costs: missed typos

A tighter request trades misses for fewer rewrites. In the 2026 study, adding a minimal-edit instruction raised the main benchmark score for six of seven models, at the cost of a 6-to-22-point drop in recall — the models fixed fewer of the real errors. Expect a cleaner reply that leaves some typos in place.

The measured result: "Imposing a minimal-edits constraint (A.2) yields the largest gain, boosting F0.5 by +1.1 to +10.9 points for all models except Claude Opus 4.6, where it is nearly neutral, but at the cost of a 6 to 22 point drop in recall."⁷ F0.5 is the field's standard score; it weights precision above recall.

The paper's minimal-edit prompt said, in part, "Make MINIMAL, PRECISE edits to fix errors. DO NOT rewrite or paraphrase."⁷ Its authors still found it underspecified. "No grammatical error categories are enumerated, so the model is left to decide what counts as an error."⁷

That is why the prompt above names the edit types. The authors' diagnosis of the declarative style: "it states a goal without saying which edits are admissible."⁷

Naming them is not a guarantee.

Their full list of 25 error categories "benefits the strongest models while remaining model-dependent overall."⁷ Some models gained; others did not.

One more finding pulls against intuition. The same paper found that batching several sentences into one request reduced edits across model families — the title is Larger Context Window, Fewer Overcorrections.⁷ Fang's team, meanwhile, saw under-correction rise with sentence length.³ Whether to paste a whole note or a paragraph at a time is not settled. Either way, the diff tells you what happened.

People run into the miss side in practice. One user on OpenAI's developer forum reported that gpt-4 caught an "except"/"expect" typo in small documents "but not for large documents."¹¹ Step 7 exists for that.

Use it as a finder, not an editor

The same 2023 data has a second finding: ChatGPT had the lowest under-correction rate of the four systems — 9.3%, 17.3% and 29.0% of sentences under-corrected, against 10.0%, 20.0% and 42.5% for Grammarly. The model is a strong finder and a loose editor. Let it point; you decide.

That reading is the fair one, and the research supports it from both ends.

Coyne's team noted that "our human raters tended to prefer the extensive rewrites."² For an essay meant for readers, the rewrite may genuinely be better. A note is different — the reader is you, later, and the words you wrote are the ones your memory is attached to.

So there is a second way to ask. Instead of a corrected note, request a list: each change as original → replacement, one per line, with the line it came from. A list cannot silently rewrite anything. Every proposed edit is a discrete pair you apply or skip, and an expanded fragment shows up as an obviously long replacement.

The list works best for short notes. For anything longer, the corrected-copy-plus-diff routine is faster — the diff does the listing for you.

Common mistakes

Most failures with AI proofreading come from one of four habits: overwriting the original, asking without naming edit types, trusting a fluent result because it reads well, and assuming a careful request caught every typo. Each one removes a check. Together they turn a correction request into an unreviewed rewrite of your own words.

• Pasting the reply over the note. Once the original is gone, thing versus matter is unrecoverable. A shorter summary is not one edit makes the same point for reduction: name the operation, then check against it.
• "Fix the typos" with nothing else. The 2026 paper's point is that a goal without a list leaves the model "to decide what counts as an error."⁷
• Reading the result instead of the diff. A fluent sentence reads as correct whether or not it is yours. Fluency is the trap in AI summaries too.
• Treating a clean diff as a clean note. A tight request misses real typos. Fewer brackets can mean fewer fixes, not fewer errors.

Each mistake is cheap to avoid — and expensive to discover later.

Frequently Asked Questions

These are the questions people type when a proofread comes back changed — several drawn from real forum threads and search phrasings. Each answer is short enough to act on, and each points back to the measured result or the routine above. None of them depends on which assistant you use.

Why does ChatGPT change my writing when I ask it to proofread?

Language models lean toward fluent text, and "proofread" leaves the scope open. In one small 2023 human evaluation, ChatGPT over-corrected 38.0% to 60.0% of sentences, depending on length, against 9.3% to 18.5% for Grammarly's free Correctness feature. A 2026 study found the same pattern across current models under a plain correction prompt.

How do I make an AI only fix spelling?

Name the edit types you allow — spelling only, if that is what you want — and forbid rewording, reordering and completing sentences. Then keep your original file and word-diff the reply against it. The named list helps on benchmarks but is not a guarantee, so the diff is the real check.

Does asking for "minimal edits" stop the AI from rewriting?

It helps, and it costs. In a 2026 study, a minimal-edit instruction improved the main score for six of seven models but cut recall by 6 to 22 points, so more real errors were left in. The authors also found a bare "minimal edits" instruction leaves the model to decide what counts as an error.

Can AI proofreading change the meaning of my notes?

Yes. A 2023 study of GPT-3.5 and GPT-4 found the models "occasionally over-edit, changing the meaning of a sentence during correction, or expanding fragments with new material." Notes are mostly fragments and shorthand, so they give a fluency editor the most to expand. Diff the result before you keep it.

Is this still a problem with newer models in 2026?

Yes, by the pattern, though no current rate is published. A paper accepted to EMNLP 2026 Findings tested GPT-5.4, Claude Opus 4.6, Gemini 3.1-Pro and four other models and found "a classic overcorrection pattern across all models" under a plain correction prompt. That was on learner-English benchmarks, not personal notes.

Why does the AI miss typos in a long document?

That is the opposite failure — under-correction. In the 2023 data, ChatGPT's under-correction rose from 9.3% of short sentences to 29.0% of long ones, and a forum user reported a missed "except"/"expect" typo in large documents only. A tighter minimal-edit request also lowers recall. Reread once after accepting changes.

A correction is the smallest edit that fixes the error. Anything larger is a rewrite — and a rewrite of your note should pass through your hands, one bracket at a time.

mnmnote.com keeps your notes as open Markdown files on your own device — so the original is one copy away, and the check is an ordinary diff.