There is a specific failure mode worth naming: a generated sentence can be correct, confident and nearly unreadable. It uses subordinate clauses where a second sentence would do, it prefers the noun form of a word to the verb form, and it reaches for "leverage" when "use" was available. None of that is an error in a grammar checker's eyes.
Readability matters more than usual when your audience reads English as a second language, because the cost of an unclear sentence lands twice — once on comprehension, once on the reader's confidence in their own comprehension. They tend to blame themselves before they blame the sentence.
The five things to look for
| Problem | What it looks like | What to do |
|---|---|---|
| Buried verb | "make a decision" · "conduct an analysis of" | "decide" · "analyse" |
| Nested clause | Two commas inside a thirty-word sentence | Split into two sentences |
| Stacked modifiers | "The new integrated cross-platform analytics dashboard" | Two sentences, or drop one modifier |
| Abstract subject | "Consideration should be given to" | "Consider" with a real subject |
| Unhedged generalisation | "Users always prefer" | Name the group or add the limit |
1. Turn nouns back into verbs
English tolerates turning almost any verb into a noun, and formal writing overuses this badly. Every conversion adds words and removes the thing doing the action.
| Written | Better |
|---|---|
| We will make a decision regarding… | We will decide whether to… |
| Conduct an evaluation of the results | Evaluate the results |
| The implementation of the change | Implementing the change |
| Facilitate communication between teams | Help teams talk to each other |
| Provides an indication of | Shows |
The last row is the clearest example: three words doing the work of one, with no extra meaning carried. Reading a draft specifically hunting for "provides an indication of" style constructions is a fast pass that removes a lot of padding.
2. Put the subject and verb together, early
English readers hold an unresolved subject in working memory while they wait for the verb. Anything inserted between the two raises the cost, and that cost falls hardest on readers processing a second language.
Hard: The upgrade, following consultation with the teams most affected and a review of the schedule implications, will proceed in November.
Clearer: The upgrade will proceed in November. We consulted the teams most affected and reviewed the schedule first.
The second version is longer in characters and shorter in effort. That trade is almost always worth taking.
3. Watch for the "which" pile-up
Generated prose loves successive clauses beginning "which" or "that". Each one appends information to the previous clause rather than making its own point, and after two of them most readers have lost the thread.
Hard: The updated policy, which took effect in March and which replaced the interim guidance issued last year, applies to all contractors working on site, which includes subcontractors engaged by third parties.
Clearer: The updated policy took effect in March, replacing last year's interim guidance. It applies to everyone working on site, including subcontractors engaged by third parties.
A quick diagnostic: if a sentence contains three clauses, it is usually two sentences pretending to be one.
4. Choose the word a person would say
Generated drafts drift toward register inflation — words that sound more professional without being more precise. This is the single most commented-on weakness in reader feedback, and there is no tool that reliably catches it, because every choice is individually defensible.
| Instead of | Write |
| utilise · leverage | use |
| facilitate | help · run |
| commence · initiate | start |
| subsequent to | after |
| in the event that | if |
| prior to | before |
| a majority of | most |
| with regard to | about |
The test is simple: would you say it out loud to a colleague? If the honest answer is no, change it.
5. Say who, and say how many
Generated text hedges toward generalities: "many teams find", "studies suggest", "users often prefer". Each of these invites the reader to wonder which teams, which studies, which users — and they usually cannot supply an answer, because there isn't one.
Vague: Many teams find that shorter meetings improve output.
Specific: Of the six teams we tested this with, five cut their weekly meeting from an hour to twenty-five minutes without anything slipping.
When you genuinely cannot name a source or figure, drop the claim rather than keeping the hedge. "Shorter meetings help" unsupported is noise; the same claim with your own numbers is information.
The ten-minute pass
- Read only the first and last sentence of each paragraph (2 min). If the topic sentence does not carry the paragraph, restructure before editing words.
- Search for buried verbs — "-ion", "-ment", "-ance" endings, and "make/give/conduct a…" (2 min).
- Split any sentence over about thirty words, or any with three clauses (3 min).
- Replace register-inflated words with the everyday equivalent (2 min).
- Find every claim without a named subject and either source it or cut it (1 min).
Do it in that order. Editing sentences before fixing structure means editing them again.
What AI does well here
Asking a tool to shorten a specific passage, or to rewrite a paragraph at a stated reading level, works genuinely well and quickly. It is also good at producing alternative phrasings when you know the current one is wrong but not why.
What it does poorly is noticing that the same comfortable construction appears nine times, or deciding that a claim is too vague to keep. Those are judgements about the whole document, and they do not appear unless you read the draft start to finish with that question in mind.
Editing time, from our 30-day benchmark
First draft: a minute. The ten-minute pass: 8–12 minutes for a 1,000-word article. In our testing across recent batches, the readability difference between a draft published as-is and one that went through this pass is noticeable even to readers who cannot say what changed.
Frequently Asked Questions
Does simpler writing look less professional?
Generally the opposite: readers associate clarity with confidence and padding with uncertainty. The exception is specialist contexts where precise terminology carries real meaning — keep those terms and simplify the sentences around them.
What sentence length should I aim for?
There is no number worth defending, but sentences over about thirty words deserve a second look, and variation matters more than average. Several long sentences in a row is what exhausts a reader.
Is this advice different for non-native readers specifically?
The principles are the same, and they matter more. Idioms, phrasal verbs and cultural references do not survive translation, so prefer literal phrasing and avoid expressions whose meaning cannot be derived from the words.
Can I just ask the tool to write more clearly?
That instruction helps but returns uneven results, because the model has no measure of what "clear" means here. Naming the specific fault — buried verbs, thirty-word sentences, inflated vocabulary — produces noticeably better output.
Should I avoid the passive voice entirely?
No. It is correct when the actor is unknown or unimportant. The problem is using it to avoid naming who did something, which is often exactly the information the reader needs.