Keep Your Voice While Fixing Your English

You paste your English draft into a grammar tool and accept every suggestion. The red underlines disappear. The text is now correct — and you read it back with a small, unplaceable feeling that somebody else wrote it.

If you write in a second language, this is probably familiar. The tools are doing exactly what they promise. The trouble is that "correct" and "yours" are two different targets, and a tool that only knows the first one will trade away the second without telling you.

Not everything that differs from a native writer's default is an error. Some of it is a choice. The skill is telling them apart.

What I can and can't speak to

Before the argument, the boundaries. I'm Jason Guo, and I build WriteFit, a bilingual (English and Chinese) writing-feedback product. That is the extent of my standing here. I am not writing from a linguist's expertise, and I am not going to claim a second-language-learner's experience I haven't told you about. What follows is a framework drawn from published research (cited, with its limits stated) and from what I have to think about when designing feedback for people who write in more than one language.

One more disclosure, because this is a product blog: WriteFit does not check grammar. It scores clarity, specificity, personal voice, strength of claim, AI-like tone, empty phrases, and reader resistance — none of which is correctness. If you need your verb tenses fixed, you need a different tool, and I will say below when that is the right call.

Three things that get lumped together

When people say "my English isn't good enough," they usually mean at least three different things at once:

1. Errors that cost comprehension. A wrong word that changes the meaning. A missing negation. A pronoun the reader cannot resolve. The reader stops, rereads, or comes away with the wrong idea.
2. Differences that cost nothing.A slightly unconventional preposition. A sentence rhythm shaped by another language. A metaphor that doesn't exist in English but lands anyway. The reader notices, and understands.
3. Sounding unlike the average. Not an error at all — just a distance from the statistical center of English prose.

Most grammar tools treat all three as one category: deviation from the expected pattern, flagged and offered a fix. That is the mechanism behind the opening feeling. It is the same force I described in the anatomy of AI voice: regression toward the mean. Voice lives in the deviations, and a tool that cannot tell cost-bearing deviations from free ones removes both.

A useful distinction from speech research

There is a research tradition that has drawn exactly this line, although for spoken language. In the 1990s, Murray Munro and Tracey Derwing studied how listeners judge second-language speech and found that three things — how strong the accent sounds, how easy the speech is to understand, and how much of the message the listener actually recovers — are related but partially independent. A speaker can have a strong accent and be perfectly easy to understand.

I want to be careful with this. It is research on speech. Nobody has shown that written "accent" works the same way, and I am borrowing the structure as an analogy — a useful way to organize your own revision — not citing it as a finding about writing. But the analogy is a good one: the question to ask of any non-standard spot in your draft is not "would a native speaker write this?" but "does this cost the reader anything?"

Why this matters beyond taste

There is a harder reason to care, and it concerns being believed.

A research team at Stanford (Liang and colleagues) tested widely used AI-text detectors on essays by non-native English writers and found that the detectors consistently misclassify non-native English writing as AI-generated, while native writing is identified accurately. The authors cautioned against using such detectors in evaluative or educational settings. I won't quote a percentage here: the exact figures vary across reports of the paper, and I have only verified the paper's qualitative conclusion.

The relevance to this post is the direction of the pressure it creates. If a second-language writer worries about being flagged, the tempting response is to make the prose smoother, safer, more generic. That instinct is understandable and, I think, backwards: the features that make your writing look machine-made to a detector — the predictable, low-variance, average-sounding sentence — are the same features a heavy-handed grammar pass produces.

A note on my own product, since I just cited this: WriteFit has an "AI-likeness" score, and it is produced by a language model following a prompt. I have not tested it for this bias, so I can't tell you it is immune to it. If you write in a second language and the score looks wrong to you, treat your own reading of the sentence as the authority, and read the feedback as one opinion rather than a verdict. I would rather say that plainly than leave you to assume otherwise. (Detectors are also not this post's main subject; I've limited them to this one section on purpose.)

A triage pass: fix the errors, keep the choices

Here is the procedure I would use. It works on a draft you have already written in your own words, and it deliberately puts grammar tools after your judgment rather than before it.

1. Write the draft without help. Not clean, not correct — just yours. If you start from a polished version, you have no way to know what you would have said.
2. Mark every spot that feels non-standard. Anything you suspect a native writer would phrase differently. Mark quickly; do not decide yet.
3. Ask the one question of each mark:would a reasonable reader stop, reread, or come away with the wrong idea? Be honest — and where you can't tell, that is the signal to ask a human reader to circle only the places where they stumbled. One external reader, asked to mark only stumbles, is worth more than ten automated suggestions.
4. Fix only the marks where the answer is yes. These are your errors in the sense that matters. This is the step where a grammar tool is genuinely the right instrument: use it on those spots, and check that the fix still sounds like you.
5. For the marks where the answer is no, decide on purpose. Keep it, and know that you are keeping it. A choice you made on purpose is voice. A deviation you never noticed is just an accident that happens to be harmless.

The last step is the one that matters most over time. The point of triage is not a cleaner single draft. It is that after a few rounds you start to feel which of your habits are carrying meaning and which are noise — and that discrimination is what a tool that fixes everything can never give you.

The triage on four sentences

An abstract procedure is easy to nod along to and hard to apply, so here are four sentences. I wrote them myself to illustrate the categories; they are not taken from any real user's draft. For each, the question is the same: does this cost the reader anything?

1. "She didn't say the budget was approved." The writer meant that she said it was not approved. As written, it only says she never claimed it was — a different claim. The reader will come away with the wrong idea. Cost: real. Fix it.
2. "We discussed about the schedule." Every reader understands it instantly, and the extra preposition changes nothing. Cost to comprehension: none.But some readers will notice it and some will judge, depending on the setting. This is the case that keeps the framework honest: the question is not only "will they understand" but also "does this setting penalize it." In a formal document you probably fix it; in a casual post, you may not bother.
3. "The numbers are small, but they speak loudly." A native writer might reach for an established phrase here. This version is understood perfectly and arguably lands harder because it is slightly off the groove. Cost: none, and possibly a gain.If you wrote it on purpose, keep it. If you didn't notice, notice it now — then decide.
4. "A product should be like water: it takes the shape of the person." The image and its rhythm may come from another rhetorical tradition, and a reader here gets it at once. Cost: none. This is exactly the kind of thing a grammar tool will tidy into something more expected, and the tidying is the loss.

Notice that the four do not sort into "errors" and "fine." They sort by what they cost, and the cost depends on the reader and the setting. That is the whole reason the question has to be asked of each case rather than answered once by a rule.

The limits of this approach

There are situations where this framework is the wrong one. In an exam, a style-enforced journal submission, or any setting where a reader will mark deviation from the standard norm regardless of whether it costs comprehension, conforming is a rational choice, and the triage above is a luxury you may not have. I am not telling you to keep what a grader will penalize.

The triage also depends on your being able to judge "would a reader stumble?" — which is exactly the thing that is hardest to judge in a language you learned later. That is why step three leans on an outside reader rather than on introspection. It is a heuristic built on an analogy from speech research and on first principles about revision, not an empirically validated writing method, and I can't show you evidence that it outperforms simply accepting a grammar tool's output.

The point

Fixing your grammar and keeping your voice are not in conflict. They only look that way when one tool is asked to do both and can see only one. The work is a division of labor: use a grammar tool where correctness is the problem, use a human reader to find where you actually stumble, and use your own judgment for the choices only you can make.

WriteFit sits in that last category by design. It does not fix your English and it does not write for you; it points at where a passage is vague, generic, or hedged, and hands the pen back so the revision stays yours. If that is the thing you are missing, read about how the training works or the method behind it, or try one session, no account needed. Bring a draft you wrote yourself — the one you are least sure is correct is a good test, because that is where a triage will teach you the most.