A Grammar Checker Caught 19 of Our 29 Planted Mistakes

Kleoniq Kosova Published October 4, 2026 Updated October 8, 2026

Photo: rawpixel, public domain (CC0)

Ask The Verbarium

The Verbarium AI

LanguageTool flagged 19 of 29 planted errors in an invented notice, while the earlier spellchecker tests flagged 8. The grammar checker missed both agreement errors, showing why a final read still matters. The test used the free service in British and American English, including its stricter setting. The stricter setting added no catches, and no false alarms appeared in the LanguageTool runs. Style checkers and the site's highlighting tool were also tested, with highlighting distinguished from detecting a mistake. Search for the remaining word confusions, inspect agreement and look up unfamiliar words. Read suggested replacements rather than trusting every explanation. The results describe one short notice on one day, not a general accuracy rate or a test of unexamined products.

The summary and the answers above were written by The Verbarium AI from this article, and can be wrong. A question you type is sent to Google to be answered. Privacy

LanguageTool flagged 19 of the 29 errors in our test notice, with no false alarms. A spellchecker alone flagged 8. Of the 21 errors that passed every spellchecker in our earlier test, LanguageTool caught 11 and let 10 through, and those 10 include both of the grammar mistakes.

This is the notice. We wrote it for the earlier test, it is 155 words long, and the 29 errors were planted on purpose.

Dear members,

Thank you for you’re patience while we recieve the the new equipment. The principle hall will be closed form Monday untill Friday, and we would of told you sooner if the dates had been definately confirmed. Please keep quite in the library, as several people has complained about noise; its a real problem and it effects everyone. Volunteers who loose there badge should collect a replacement at the front desk, then walk passed the office to seperate the recycling. The trip to london is to far for the casual link between weather and attendance to be tested, but the organisers recognized that the colour of the posters and the color of the badges should match. Lunch will be served in the refractory, and the english class was lead by more then twenty volunteers. Their will be a meeting at 6pm; these meeting is informal and it’s outcomes will go on the notice bored.

What was tested, and what was not

All of this was run on 4 October 2026, on the notice exactly as printed above.

LanguageTool, through its free public API, once with the language set to British English (en-GB) and once with American English (en-US). The API reported version 6.9-SNAPSHOT, built 1 September 2026. We then repeated both runs with the stricter level=picky setting. That is four requests in all.
proselint 0.16.0 and write-good 1.0.8, two open-source style checkers that run on your own computer, each with its default settings.
Our own checker from the earlier article, which highlights sound-alike words for you to look at.

We could not run Microsoft Word, Google Docs or Grammarly from the machine we test on, so they were not tested and this page says nothing about what they would catch. LanguageTool’s paid Premium version was not tested either. Each API response carried the note “You might be missing errors only the Premium version can find.”

Scoring follows the earlier test. Each planted error is either flagged or not. The two British and American mixes count as caught if either half is flagged. A flag on anything that is not one of the 29 is counted separately.

All 29 errors against each tool

Of the 29 planted errors, the spellcheckers flagged 8, LanguageTool 19, proselint 3 and write-good 1. Our own checker highlighted 18, which is not the same thing as catching them: it marks every sound-alike for you to look at, right or wrong.

The full table: every error, and which tool flagged it

In the notice	Should be	Spellcheck	LanguageTool	proselint	write-good	Our checker
you’re	your	no	yes	no	no	yes
recieve	receive	yes	yes	no	no	no
the the	the	no	yes	yes	yes	yes
principle	principal	no	no	no	no	yes
form	from	no	no	no	no	yes
untill	until	yes	yes	no	no	no
would of	would have	no	yes	yes	no	yes
definately	definitely	yes	yes	no	no	no
quite	quiet	no	no	no	flagged, for another reason	yes
several people has	have	no	no	no	no	no
its	it’s	no	yes	no	no	yes
effects	affects	no	yes	no	no	yes
loose	lose	no	no	no	no	yes
there	their	no	no	no	no	yes
passed	past	no	yes	no	no	yes
seperate	separate	yes	yes	no	no	no
london	London	yes	yes	no	no	no
to (is to far)	too	no	yes	no	no	yes
casual	causal	no	no	no	no	yes
organisers with recognized	one variety	yes	yes	no	no	no
colour with color	one variety	yes	yes	yes	no	no
refractory	refectory	no	no	no	no	no
english	English	yes	yes	no	no	no
lead	led	no	yes	no	no	yes
then (more then)	than	no	yes	no	no	yes
Their	There	no	no	no	no	yes
these meeting	this meeting	no	no	no	no	no
it’s	its	no	yes	no	no	yes
bored	board	no	yes	no	no	yes
Total	29	8	19	3	1	18

The spellcheck figure is the earlier result: an error counts if at least one of the four Hunspell and Aspell runs flagged it. LanguageTool gets one figure because its British and American runs caught the same 19 errors, and the picky setting added nothing to either.

Bar chart of planted errors flagged out of 29: spellcheckers 8, LanguageTool 19, proselint 3, write-good 1 — The totals for the four tools that flag errors.

LanguageTool by kind of error

LanguageTool caught everything the spellcheckers did: the four misspellings, the two lower-case proper nouns and the two British and American mixes. On top of that it caught 8 of the 12 wrong sound-alikes, 1 of the 4 one-letter slips that make a real word, the doubled word and “would of”. It missed four sound-alikes (principle, loose, there, Their), three one-letter slips (form, quite, casual), both agreement errors and the rare word refractory.

Chart comparing a spellchecker and LanguageTool on nine kinds of error: LanguageTool adds eight of the twelve sound-alikes, and neither catches the agreement errors — Where LanguageTool adds to a spellchecker, and where neither helps.

The eight sound-alikes it caught are the ones where the next word gives the mistake away: you’re patience, its a, it effects, walk passed, is to far, was lead by, it’s outcomes, notice bored. It also caught more then twenty. The ones it missed sit in sentences that still look grammatical to a rule: the principle hall, closed form Monday, keep quite, the casual link.

The agreement errors are the result we did not expect. A grammar checker passed “several people has complained” and “these meeting is informal”, in both varieties and at both settings. Of everything in this test it is the finding I would remember, because agreement is exactly the kind of mistake a grammar checker is for.

Does the language setting matter?

Only for the two spelling mixes, and both runs caught both. Set to American English, LanguageTool flagged organisers and colour and said each “is British English”. Set to British English, it flagged color as “American English” and asked whether we would like the Oxford spelling organizers, adding that organisers “is also correct”. Accepting both British suggestions leaves organizers, recognized and colour twice, which is consistent. The spellcheckers in the earlier test flagged these words as plain misspellings and never named the variety. British and American spelling differences has a checker for the mix alone.

Right fix, wrong explanation

LanguageTool’s first suggestion was the correct word at all 19 flags. Two of its explanations were off even so. For effects it suggested affects and glossed it as “experiences of emotion”, which is the noun, when the sentence needs the verb. For bored it suggested board, glossed as “a committee”, when the notice means a noticeboard. Read the suggestion, not the gloss.

write-good flagged quite, but as a “weasel word”, with no hint that quiet was meant. We did not score that as a catch. proselint’s fix for “would of” was would’ve, which is right but informal for a notice, and for colour and color it reported the inconsistency without choosing a side.

Flags on things that were not errors

LanguageTool, all four runs: nothing.
proselint: 6pm, twice. It wanted periods in “p.m.” and a space before it.
write-good: be closed, be tested and be served as “may be passive voice”, and several as a “weasel word”.
Our checker: three correct uses of to, one then and weather. It gave no message, because it highlights every sound-alike by design.

proselint and write-good are style checkers and it shows. Between them they found the doubled word, “would of” and one spelling mix. They are not a substitute for a proofread and do not claim to be.

What is still left for you

After a spellchecker and LanguageTool, ten errors remained in the notice. They fall into two groups.

Seven you can search for. principle, form, quite, loose, there, casual and Their. Each is on the search list in the spellcheck article, with a test to apply at every hit. So the list of searches gets shorter with a grammar checker, but it doesn’t go away: principal and principle, from and form, quiet and quite, lose and loose, their and there, causal and casual still need your eye.

Three that only reading finds. “Several people has”, “these meeting is” and refractory for refectory. No tool on this page flagged any of them. The slow read aloud described there is still the step that catches agreement, and a rare word still has to be looked up.

Does our checker add anything on top?

It highlights all seven of the searchable ones in one pass and none of the other three. That is the same ground the searches cover, done faster. The cost is noise. Once LanguageTool’s 19 fixes were applied to the notice, our checker made 21 highlights, and 14 of them were words now used correctly. The example below is most of that corrected notice: 18 highlights, of which 7 are the errors LanguageTool left.

Check your text for errors spellcheck cannot see

Paste your text here

Runs in your browser. Your text is not sent anywhere or stored.

Our checker's result on most of the corrected notice, with eighteen sound-alike words highlighted — Our checker on most of the corrected notice: 18 highlights, 7 of them errors that LanguageTool left.

Before you paste work text into it

LanguageTool checks your text on its own servers, so whatever you paste is sent to them. Its privacy policy says the company collects content you input, and that it “does not use inputs from any LanguageTool user (Individual or Team Plan) to train its language algorithms or models”. Its documentation says the version running on the free service “may be the latest official release or some snapshot”, so results can change without notice, and that you can set up your own instance if the text is confidential. proselint, write-good and our checker run on your own machine or in your browser and send nothing.

One limit on all of the above: this is one short notice with errors we chose, run on one day. It tells you which kinds of mistake got through, not how often they will in your own writing.

Keep reading

Usage and SpellingAugust 11, 2026