“The English requires editing by a native speaker” is one of the most common lines in a reviewer report, and it is almost never about accent. Reviewers rarely quote the sentence that stopped them. They read a paragraph twice, lose the thread, and convert that friction into a verdict about the language. The manuscript that produced it usually passed a grammar checker with nothing flagged.
That is expected behavior, not a failure of the checker. Grammar tools work inside a sentence, on strings, against rules about agreement, punctuation, and usage. They have no model of your study design, so “we demonstrate that X causes Y” over correlational data parses as clean English. They do not track what a term meant forty pages earlier, so a construct that starts as “adherence,” becomes “compliance” in the methods, and returns as “treatment uptake” in the discussion raises no flag. They cannot count how many nouns a “this” could refer to.
Line-level editing sits between a spell checker and a developmental edit. It works on the prose sentence by sentence without touching your argument, and the passes below are the ones that change what a reviewer experiences.
What does a grammar checker actually miss?
It misses everything that depends on context outside the sentence. A useful language pass works in four layers, and only the first is mechanical.
- Mechanics. Grammar, punctuation, spelling, and usage under your target style guide (AMA, APA, Chicago, or the journal’s house style).
- Clarity. Ambiguous pronouns, agentless passives, sentences whose subject and verb are separated by twenty words of qualification. These cost the reader effort that they will charge back to you.
- Cohesion. One label per construct, abbreviations defined where they are first used and again in the abstract, tense that stays consistent within a section.
- Field conventions. Claim verbs matched to the study design, hedging calibrated to the evidence, nomenclature and statistical reporting in the form your discipline expects.
Layers two through four are why a manuscript can be mechanically clean and still read like a draft. They also account for most of what a copyeditor at a journal will not fix for you, because by the time copyediting happens, the reviewers have already formed their opinion.
How do you find ambiguity you cannot see?
You find it by searching for the words that hide it, because you cannot read for it directly. You know what you meant, so your eye supplies the missing referent every time.
Run the referent test. Search your manuscript for “this,” “these,” “it,” “they,” “the former,” “the latter,” and “respectively.” For each hit, apply one rule: every demonstrative gets a noun after it. Not “This suggests a dose effect” but “This dissociation suggests a dose effect.” Then check the preceding sentence for how many nouns could plausibly fill that slot. If there is more than one, the reader is guessing, and half of them will guess wrong.
The second search is for agentless passives in the methods: “samples were excluded,” “outliers were removed,” “the threshold was set.” Each one drops the two facts a reader needs to reproduce the work: who decided, and by what criterion. Rewriting to “we excluded 14 samples with RIN below 7” adds an actor and a number in the same move.
The third test needs another person and takes ten minutes. Read your abstract aloud to a colleague outside your subfield, then ask them to state what you did and what you found. Where they hesitate, the sentence is ambiguous. Where they restate something you did not claim, the sentence is worse than ambiguous.
Does your terminology hold from introduction to discussion?
Usually not, because varying your word choice is a virtue in general English prose and a defect in scientific prose. Every synonym you introduce for elegance asks the reader to decide whether you mean the same thing. Pick one label per construct, per measure, and per group, and use it in the title, abstract, methods, results, figures, and discussion without variation.
The census is mechanical. Export the manuscript to plain text and count what you actually wrote:
tr '[:upper:]' '[:lower:]' < manuscript.txt \
| grep -oE "[a-z][a-z-]{3,}" \
| sort | uniq -c | sort -rn | head -60Read the list for near-duplicates: participants and subjects, cohort and sample, method and approach, accuracy and performance. For each cluster, decide which one is the term, then replace the others. The same list exposes the opposite problem, one word doing two jobs, when “model” covers both your statistical model and the animal model.
Abbreviations need their own pass, and the rule is stricter than most authors assume. The abstract is indexed and read on its own, so an abbreviation used there has to be defined there, and defined again at its first use in the body. Search each abbreviation and check three things: it is expanded at first use in each of those places, it is used at least three more times (otherwise cut it and write the words), and it never appears in the title.
Tense follows the same logic. Methods and results are past tense because they describe what you did and observed. What a figure shows is present tense, because the figure keeps showing it. Established knowledge is present tense. Drift between these inside one paragraph is one of the most frequent things a language editor changes, and one of the least likely to be caught by a rule-based checker.
Is your hedging calibrated to the evidence?
Calibration means the claim verb matches the design that produced it. Observational data supports “is associated with,” “predicts,” and “is consistent with.” It does not support “demonstrates,” “establishes,” or “shows that X causes Y,” and a reviewer who works in your field reads that mismatch as either carelessness or salesmanship. Neither helps you.
Build the claim table. Copy every sentence from your abstract and discussion that contains a claim verb into two columns: the verb you used, and the design that would justify it. Randomized and prespecified supports causal language. Correlational supports associational language. Post hoc supports “exploratory” and nothing stronger. Any row where column two does not license column one gets downgraded, and the downgrade is usually one word.
Statistical language is where this goes wrong most visibly. The American Statistical Association’s statement on p-values is explicit that a p-value measures neither the probability that the hypothesis is true nor the size of an effect, which rules out “highly significant” as a description of magnitude and rules out “no effect” as a description of p above 0.05. It also rules out the phrase reviewers enjoy quoting back: a result that “approached significance” or showed “a trend toward significance” was not significant under the threshold you chose, and describing it otherwise invites a comment about analytical flexibility.
Miscalibration runs in the other direction too, and it is the more common failure among authors writing in a second language. Stacked hedges (“these results may possibly suggest a potential association”) read as no claim at all, and a reviewer cannot evaluate a claim you did not make. One hedge per claim, chosen deliberately: “may,” “suggests,” or “is consistent with,” not all three.
Which conventions belong to your field rather than to English?
Most of them, and this is the layer where general-purpose editing help does the most damage. What a non-specialist editor corrects as awkward is frequently a discipline requirement.
The recurring ones are worth checking directly. Species names are italicized, with the genus abbreviated after first use. Human gene symbols are uppercase and italic while the protein products are not, per the HGNC guidelines, and the equivalent rules differ for mouse, yeast, and Drosophila. SI units take a space between the number and the unit, with the exception of the degree sign for angles. Statistical reporting follows your field’s reporting guideline rather than your preference, and the EQUATOR Network hosts the checklist for your study type: CONSORT for trials, PRISMA for systematic reviews, STROBE for observational studies, ARRIVE for animal research.
The fastest way to align is to work from evidence rather than memory. Open the three most recent papers you cited from your target journal and note how they handle five things: tense in the methods, how effect sizes and confidence intervals are reported, whether p-values are given exactly or as thresholds, how the first person is used, and how figures are referenced in the text. That is your local convention. It beats any general style advice, including this post’s.
When should the language pass run?
After the argument and structure are stable, and before you reformat for the journal. If sections are still moving, every pass regenerates the same inconsistencies: a term you standardized reappears in text you pasted from an older draft, and an abbreviation you defined in section 3 now first appears in section 5.
A workable sequence:
- Draft for content and structure.
- Get a review of argument, methods, and contribution, and revise for substance.
- Run the line-level language pass on the revised manuscript.
- Format to the journal’s guidelines, then reread the abstract and figure captions, which formatting most often breaks.
Then apply the freeze rule: after the language pass, any paragraph you add gets its own pass before submission. Text written in the last hurried days, typically a new limitations paragraph or a response to a co-author’s comment, is where the sentence a reviewer quotes usually comes from. It is the only text in the manuscript that no one has read slowly.
What do editors expect you to disclose?
Language help, on its own, is standard and generally needs no disclosure. Journals sell it themselves, and using a professional editing service is not an ethics question. AI assistance is treated differently: the ICMJE recommendations and the COPE position statement on AI tools both hold that such tools cannot be authors and that authors remain responsible for every word, with disclosure expected when the tool contributed to the writing. Check your target journal’s instructions, because thresholds for what must be disclosed still vary between publishers.
Responsibility has a practical form here. Any tool that rewrites a sentence can also change what it claims, and the changes that matter are small: an “associated with” that becomes “leads to,” a value that loses a decimal, a gene symbol that gets autocorrected into a word. After accepting suggestions, diff the edited manuscript against the previous version and read only the numbers, units, nomenclature, and claim verbs. This takes a few minutes and catches the class of error that reviewers treat as a data integrity problem rather than a language one.
Dedicated academic tools narrow the field-convention gap somewhat, which is why they exist alongside general grammar checkers. Fukurō’s Language check runs mechanics, clarity, cohesion, and rigor passes over the full manuscript and returns tiered suggestions you accept or dismiss individually, with the diff kept visible for exactly the reason above. It does not know whether your claim is true, it has no access to your data, and it does not replace a co-author who knows the literature. If you are weighing options at this stage, our criteria-based comparison of AI review tools covers what each type of tool can and cannot check.
None of these passes are visible in the finished paper. That is the point. The reviewer will never know which one caught the ambiguous pronoun or the overstated verb. They will only notice whether they had to read any sentence twice.