Is using ChatGPT for grammar cheating? At almost every university, no, and one of the ten policies reviewed for this article says so in writing. The exceptions are narrow, and they turn on what the tool does to your text rather than on why you used it.
Why Grammar Help Was Never the Problem
Spelling and grammar correction has been built into word processors for thirty years, and no policy examined here bans it. Oxford is the most direct: its guidance permits "local editing tools" such as "grammar assistants" and "spell-checkers" and states they "need not be declared". The reasoning given is that these tools "only make small, local changes", usually affecting a few words at a time.
UCL reaches the same place by a different route, permitting a "Language and Writing review" covering "structure, fluency, presentation, grammar, spelling, punctuation, and language translation". Harvard's FAS guidance is more candid still, admitting that nobody currently thinks twice about spell-checking and grammar-checking tools in word processors.
So the anxiety driving this search is mostly misplaced. What it is picking up on, correctly, is that ChatGPT does not behave like a spellchecker.
Is Using ChatGPT for Grammar Cheating? The 3 Cases Where It Becomes That
Your language is the thing being assessed. UCL permits translation help but withdraws that permission where the student is being assessed on language ability itself. On a language module, or an assignment marked partly on written expression, correction tools remove the thing being measured.
The edits stop being local. This is the common one. Every permissive policy quoted above contains a ceiling. Oxford's covers "small, local changes". UCL's becomes misconduct "if substantive changes to content have been made by the reviewer or software or at their recommendation". Edinburgh permits direct editing "only to indicate and correct minor issues", meaning items "localised, typically single words or punctuation items within existing sentences", and forbids a proofreader from rewriting text or altering "the sequence of information or line of argument".
Disclosure was required and you skipped it. Cambridge's rule is that "a student using any unacknowledged content generated by artificial intelligence within a summative assessment as though it is their own work constitutes academic misconduct, unless explicitly stated otherwise in the assessment brief". Monash and Melbourne both require an acknowledgement where AI has been used. None of these is triggered by fixing a comma, but all of them are triggered by generated sentences arriving in your submission.
The Difference Between Suggesting and Rewriting
This is the part that actually matters, and it is mechanical rather than moral.
A conventional grammar checker underlines a problem and offers a change. You accept or reject each one, and every sentence in the finished document is a sentence you approved individually. ChatGPT, given the same paragraph and a request to fix the grammar, hands back a rewritten paragraph. Some of those sentences are now the model's, and you never made a decision about them one at a time.
Your intent was identical in both cases. The output is different in kind. Edinburgh encodes exactly this distinction by requiring proofreader edits to be tracked, "e.g. using 'Review > Track Changes' in Microsoft Word", so that each change remains visible and reversible.
The practical version: paste back a paragraph you cannot account for sentence by sentence, and you have moved from correction to authorship. If a passage does need real rewriting, doing that with a human rewriter tool and reviewing each version keeps the decisions yours.
If English Isn't Your First Language
Worth saying plainly: none of the twelve institutions checked for this article publishes an explicit carve-out for second-language writers. UCL comes closest by permitting language translation, then withdraws it where language is being assessed. So the honest advice is not that you have an exemption, because no policy found grants one. It is that language-level correction sits on the permitted rungs anyway, for everyone. Where phrasing needs more than a comma fix, an AI to human rewriter worked over your own sentences keeps it on those rungs.
There is a second problem that lands disproportionately on this group, and it has nothing to do with policy.
Does Grammar-Only Editing Trip a Detector?
Detectors are unreliable in a way that is specifically unfair to non-native writers. Liang and colleagues (Patterns, 2023) ran seven widely-used GPT detectors over 91 human-written TOEFL essays and found an average false positive rate of 61.22%. Every one of the seven flagged 18 of those essays, and 89 of the 91 were flagged by at least one detector. Accuracy on US eighth-grade essays was near perfect by comparison.
The mechanism is perplexity, meaning how predictable each word choice is. Simpler, more conventional phrasing scores as more predictable, and predictability is what these tools read as machine-written. Writing carefully in a second language produces exactly that pattern.
Whether grammar-only correction moves a score in one direction or the other is not something the published evidence settles, and this post will not guess at it. What is documented is that the scores are noisy enough that Monash refuses to let them be used as evidence at all. If you want the underlying mechanics, how detectors reach a score is covered separately, as is the stricter set of rules that applies to a dissertation.