Somebody on your team is already pasting customer emails, proposals and documentation into a chatbot with the instruction "fix my grammar". It is fast and nobody has a human proofreader for every text. The problem is that "reads better" is not the same as "only the errors were fixed", and a changed name or number is hard to spot once the text reads smoothly.
This article gives a one-page rule: where the text may go, how to ask for corrections only, how to check every change mechanically, and when a human editor is needed.
Why "fix my grammar" is the wrong request
Language models tend to improve text rather than merely correct it. A 2023 evaluation of ChatGPT on grammatical error correction by Fang and colleagues reports that it corrects errors so freely that the sentences become very fluent, "possibly due to its over-correction tendencies and not adhering to the principle of minimal edits". The same paper reports that ChatGPT could not effectively correct some errors that span several sentences, such as agreement, coreference and tense.
Coyne and colleagues compared GPT-3.5 and GPT-4 corrections with human reference corrections and observed differences in editing strategies. Davis and colleagues tested seven open-source and three commercial models on English learner text. They found that commercial models do better on benchmarks annotated with fluency corrections than on benchmarks annotated with minimal edits, and that several open-source models outperformed the commercial ones on minimal-edit benchmarks.
These are benchmarks on learner text, not your proposals. Our judgment is that the direction still matters: the tool is rewarded for fluent output, while you want the smallest safe change. In one small test, the Czech site josefd.cz reports that a plain request to fix spelling and grammar changed a colleague's surname and rewrote a verbatim quotation. That is one run, not a measurement, but it is the failure that costs a company credibility.
Treat an AI proofreader as a source of suggestions, never as an editor who may change meaning.
Decide where the text may go
Before the prompt, decide where the text may go. That depends on the tool and the plan, and the vendors' own pages are the only reliable source. Three vendor examples:
- Claude (consumer plans). Anthropic says it may use chats to improve its models if you allow it, or if a conversation is flagged for safety review; Incognito chats are not used to improve Claude. Anthropic says deleted conversations leave its back-end storage within 30 days, and data used for training may be kept de-identified for up to five years, see its retention page.
- Gemini apps. Google says that temporary chats and chats with Keep Activity off are retained for 72 hours. Temporary chats are not used to train its models; with Keep Activity off, Google says it does not use future chats to improve its AI models if you do not submit feedback. With Keep Activity on, Google asks users not to enter confidential information they would not want a reviewer to see, and says reviewed chats are retained for up to three years.
- Grammarly. For individual accounts, Grammarly states that Product Improvement and Training is on by default. On multi-user Pro accounts only an administrator can change it.
We did not check every provider, and these pages change. Check the current page for each tool you use and review it each quarter. We suggest three tiers (our recommendation, not a vendor rule):
- Company-approved tool and plan. Normal business text, once your IT lead has read the terms for that plan.
- Consumer chatbot in a no-training mode. Only text you would accept being stored for days, in a temporary or incognito chat.
- Placeholders. Replace names, prices and confidential details with markers such as [CLIENT] before the text leaves your systems, and restore them afterwards.
The question is not whether AI proofreading is allowed, but which tool, which plan and which text.
Ask for corrections only, and ask for a list
A strict request narrows what the model may do. This template is our own suggestion and has not been validated by any vendor:
Correct only spelling, grammar and punctuation errors in the text below.
Do not rephrase, shorten, reorder or improve the style.
Leave names, numbers, dates, prices, product names and quotations exactly as written.
Return (1) the corrected text and (2) a list of every change as "original -> corrected".
If you are unsure whether something is an error, do not change it; list it
separately under "Uncertain".
The "Uncertain" list matters: it shows the places a person should decide, such as a product name that looks like a typo or a verbatim quotation that looks ungrammatical.
Expect a trade-off: a strict request may leave a few real errors in place, which is the right default where an unfixed comma costs less than an altered figure. Keep the prompt in one shared place.
A strict prompt reduces rewriting, but it does not prove that nothing else changed.
Check every change with a diff
The change list a chatbot returns is its own description of what it did, and it may not match the text it returned. The only check that does not depend on the model is a mechanical comparison of the original and the corrected text. Use what your team already has:
- Microsoft Word. The legal blackline option compares two documents and shows the differences in a new third document, leaving the originals unchanged, at word or character level.
- Google Docs. Compare documents creates a third document that shows the differences as suggested edits.
- Plain text, Markdown or code.
git diff --word-diffshows changes word by word, marking removed and added words. Save both versions as files and compare them.
Then apply one rule. Every difference in the comparison must appear in the model's list, and you must accept each one. If the comparison shows a change the list did not mention, reject the whole run and start from the original. Check any figure, name or quotation in the comparison against the source yourself.
A comparison catches changes, not missed errors. Because of the cross-sentence weakness reported above, also read the corrected text once as a whole.
Trust the comparison, not the description of the comparison.
When the text belongs to a human editor
Some text should not rely on AI proofreading alone. Our recommendations:
- Contracts, terms and legal notices, where one changed word can change an obligation.
- Print and published material, which cannot be corrected after distribution.
- Text bound to approved terminology, such as regulated product descriptions, safety instructions or technical standards.
- Languages or domains where the tool is weak. Test the tool on a sample of your own text with known errors first; in our judgment that tells you more than any benchmark.
For routine emails and internal documentation, the process above is usually enough.
The higher the cost of a single wrong word, the more the final check should be a person.
A one-page rule to adopt this week
- List approved tools and plans and what each may receive (the three tiers).
- Replace confidential passages with placeholders.
- Use the shared strict template and require the change list and the "Uncertain" list.
- Compare original and corrected text with Word, Google Docs or
git diff --word-diff, and reject any unlisted change. - Send contracts, print and terminology-bound text to a human editor.
Add these five points as a clause in your AI policy; our one-page acceptable use policy for employees is a good starting point. The same check-against-the-original logic applies to summarising long documents with AI, and drafting rather than correcting is covered in AI copywriting.
If you want help turning this into a working process for your team, with approved tools and checks that fit your documents, talk to our AI consulting team.
Sources
- Fang, T. et al., 2023, "Is ChatGPT a Highly Fluent Grammatical Error Correction System? A Comprehensive Evaluation", arXiv:2304.01746. Cited for the authors' over-correction finding.
- Coyne, S. et al., 2023, "Analyzing the Performance of GPT-3.5 and GPT-4 in Grammatical Error Correction", arXiv:2303.14342. Cited only for editing-strategy differences.
- Davis, C. et al., 2024, "Prompting open-source and commercial language models for grammatical error correction of English learner text", arXiv:2401.07702.
- Anthropic, 2026, Is my data used for model training? (page dated 16 March 2026) and How long do you store my data? (page dated 1 July 2026). Vendor statements, consumer plans.
- Google, 2026, Gemini Apps Privacy Hub (page dated 26 October 2026). Vendor statements.
- Grammarly, 2026, Product Improvement and Training Control. Vendor statement.
- Microsoft, 2026, Compare document differences using the legal blackline option.
- Google, 2026, Compare documents in Google Docs.
- Git project, 2026, git-diff documentation, option
--word-diff. - josefd.cz, 2026, AI korektura českého textu (in Czech, 4 November 2026). One run, cited as an example only.
