Researchers Just Fixed the Most Annoying Habit of AI Grammar Checkers

If you have ever let an AI polish an email, you know the feeling: you write something perfectly fine, and the software insists on rewriting it into corporate mush. That habit has a name in the research world — overcorrection — and it is the single biggest reason people give up on AI proofreading tools.

A new paper from researchers Kateryna Karpo and Artem Chernodub tackles that problem directly. Their approach is prompt-based, which simply means they give the AI written instructions rather than retraining it. First, they hand the model a detailed checklist of what actually counts as a grammar mistake, so it knows exactly where its job starts and stops. Second, and more surprisingly, they feed the AI several sentences at once instead of one at a time. The researchers suspect that when the model juggles more text, it loses the confidence to meddle with sentences that were never broken.

Finally, they used another AI to help sharpen the wording of those instructions. Powered by Google's Gemini 3.1-Pro, the finished prompt scored 78.32 on BEA-2019, a widely used grammar-correction test. That puts it just 0.38 points behind the best system trained specifically for the task — a gap that used to require serious computing budgets to close.

Why does this matter outside the lab? Because the same pattern shows up everywhere AI touches writing. Chatbots that "improve" your message, translation tools that smooth out your tone, assistants that edit your reports — all of them struggle with the same instinct: doing too much. The fix here is cheap and reusable. It means better writing tools for everyone, and less of your voice quietly disappearing in the edit.