I once sent a contract to a client with “defiantly” where I meant “definitely.” Twice, in the same document. Spellcheck shrugged. The client noticed inside an hour. That small humiliation is exactly the gap an AI grammar checker is built to close: not just flagging words that don’t exist, but spotting real words sitting in the wrong place.
There are now dozens of these tools, and they are not interchangeable. Some are rule engines with a marketing budget. Others are language models that read a whole sentence before deciding whether “which” should be “that.” Knowing the difference saves you from accepting a suggestion that quietly makes your writing worse.
What these tools actually do under the hood
Older checkers worked from thousands of hand-written rules plus a dictionary. They caught “your” versus “you’re” reliably and fell apart the moment a sentence got complicated. Modern tools run on models trained on enormous volumes of text, which means they judge a word by the company it keeps rather than by a lookup table.
That shift explains both the wins and the weirdness. A model-based checker can tell that “The team are reviewing the proposal” reads fine in a Manchester office and oddly in a Chicago one. It can flag a comma splice buried in a 40-word sentence where a rule engine would have given up. It can also, occasionally, insist that a perfectly good technical term is a typo because it has never seen your database column names before.
The fixes you’ll notice in the first week
- Agreement and tense slips: subject-verb mismatches, stray past-tense verbs inside present-tense paragraphs.
- Punctuation in long sentences: comma splices, missing serial commas, dialogue punctuation.
- Word choice: affect/effect, complement/compliment, elicit/illicit, and the less/fewer distinction in countable lists.
- Repetition: the same verb opening four consecutive sentences.
- Register: “Hey” in an email to a regulator, “pursuant to” in a Slack message.
Most people find their error rate drops sharply in the first fortnight, then plateaus. The plateau is where the real work starts.
Where AI checkers get it wrong
False positives on specialised language
If you write about APIs, medical dosages, or legal citations, expect a steady stream of red underlines that are not errors. Product names, code snippets, and deliberate lowercase brand styling will all trigger suggestions. Build a personal dictionary early and add terms as they appear, or you’ll spend more time dismissing flags than writing.
The passive voice crusade
Many tools were trained partly on style guides that treat passive constructions as a moral failing. “The samples were destroyed in transit” is passive and also the correct sentence, because nobody knows who destroyed them. Accept the suggestion and you get “Someone destroyed the samples in transit,” which invents a fact.
Flattening your voice
This is the subtle one. A checker tuned for maximum clarity will sand down idioms, sentence fragments used for rhythm, and regional phrasing. Two passes later, your draft reads like everyone else’s. That risk is identical when you use rewriting tools, which is why it pays to read up on using an AI paraphrasing tool without losing your voice before you accept a hundred suggestions in a row.
Non-native speakers get over-corrected
Studies of automated writing feedback have consistently found that learners accept far too many suggestions without reading them. A tool that turns “I have been working here since three years” into idiomatic English is genuinely helpful. One that also rewrites every slightly unusual but perfectly grammatical structure teaches the wrong lesson. Reading each change before accepting it is worth more than any premium tier.
Features that actually earn their keep
Feature lists are long and mostly identical, so here is what separates a tool you keep from one you cancel in month two.
- Where it runs. A browser extension, a desktop app, and a Google Docs add-on together cover roughly 95% of real writing. A tool that only lives in a web editor gets forgotten.
- Custom style rules. Enforcing things like “no exclamation marks in customer emails” or a house spelling of “organisation” beats a hundred extra synonyms.
- Data handling. Check whether your drafts train the model and whether an enterprise tier offers a no-training clause. If you handle client contracts, this is not a detail.
- Calibrated tone detection. A confidence score you can adjust is better than a red “sounds aggressive” banner on a sentence that reads perfectly neutral.
- Offline mode. Rare, increasingly available, and useful on flights or locked-down networks.
A three-pass workflow that beats accepting everything
Pass one: structure, checker off
Write the draft, then read it aloud or run your system’s text-to-speech. Grammar tools are useless at telling you that paragraph four repeats paragraph two. Fix the argument first, in your own words.
Pass two: line editing, checker on
Now run the tool. Work sentence by sentence and sort each suggestion into accept, reject, or rewrite-it-yourself. The third bucket is where the learning happens: the checker spots a clumsy phrase, you write something better than its suggestion.
Pass three: the read that catches what the machine missed
Read the final version backwards, sentence by sentence. It feels ridiculous and it works, because you stop tracking meaning and start seeing words.
Do you need to pay?
For casual writing, no. Plenty of AI tools you can use without signing up will handle a cover letter or a short report in under ten minutes. Free tiers usually cap word counts and hide tone and clarity suggestions behind a paywall, which is a fair trade at a few hundred words a week.
Paid plans run from about $12 to $30 a month. Easy to justify if you send client-facing documents daily and one typo costs you credibility. Before your card comes out, read a candid breakdown of what Grammarly’s AI does well and where it falls short, because the most popular tool is not automatically right for academic citations or for code-adjacent writing.
What no grammar checker can do
These tools check language, not truth. They will happily polish a paragraph built on a statistic you half-remembered from a 2019 blog post. If your work involves factual claims, bolt on a separate verification step: open the source, confirm the number, confirm the date. The same discipline applies to any generative system, which is why serious teams spend so much effort on grounding and on building AI systems that stay truthful to their sources instead of trusting fluent output.
They are also poor judges of whether a document is the right length, pitched at the right reader, or answering the question that was actually asked. Those are structural calls that need a human with context.
Build a personal error profile instead of chasing zero mistakes
Here is the habit that separates people who improve from people who just run spellcheck. Keep a plain text file called my-errors.txt and add a line every time a checker catches the same class of mistake. Three weeks in, you’ll have a list of eight or ten recurring items. Mine runs to comma splices, “it’s” for “its” under deadline pressure, and starting sentences with “So.”
Search your drafts for those specific patterns before you open any tool. Targeted proofreading takes about ninety seconds and beats automated scanning every time. The checker then does the job it is actually good at: catching the errors you don’t yet know you make, not standing in for knowledge of your own writing. To see where it sits beside research, outlining, and visual tools, this breakdown of a practical AI stack for content creators shows how a grammar checker slots in without swallowing the whole workflow.