How Does Grammarly Work for Students? A Teacher's Guide
A student hands in a paragraph that came back from Grammarly looking suddenly polished — every comma placed, the passive voice flagged and fixed, the wandering sentence about theme somehow tightened into two clean lines. You read it. You genuinely cannot tell what the tool changed, what the student decided, or whether either of them understood why any particular edit was an improvement. That moment is not a cheating problem. It is a literacy gap — and no one thought to brief you on what the machine was actually doing. That’s not on you. Once you can see the three-layer system running under the green, blue, and red underlines, you can teach it.
Grammarly works on three stacked layers: a dictionary spell-checker, a grammar layer that parses sentence structure against rules plus a machine-learning model trained on error data, and a style/tone layer that rewrites whole sentences. The tool shows high precision — about 91% of what it flags is genuinely flagged correctly — but low recall, missing roughly half of real errors (F1 score ≈ 0.66 in a 2025 peer-reviewed study). It over-flags passive voice, marks a majority of “correctness” suggestions as optional rather than actual errors, and the style layer nudges writing toward Standard American English in ways that are not neutral. Teaching students how it works and where it fails turns a shortcut into a thinking tool.
How does Grammarly work, step by step?

Grammarly runs three distinct systems in sequence. Students see the output as colored underlines — red, blue, green — but the decision-making behind each color is different in kind, not just degree.
| Layer | What it checks | How it decides | Underline color |
|---|---|---|---|
| Spell-check | Is this a real word? | Dictionary lookup + edit distance (find the closest real word) | Red |
| Grammar-check | Is this sentence grammatically well-formed? | Tokenization + parsing + a machine-learning model trained on grammatical error correction (GEC) data | Blue |
| Style / tone | Could this be clearer, more concise, more direct? | An LLM-era model that rewrites whole clauses and sentences | Green |
The spell-check layer is the oldest and simplest. It breaks the text into tokens — roughly word-sized chunks — and checks each against a dictionary. When a token is not found, it calculates edit distance (how many character substitutions get you to the nearest real word) and suggests the closest match. See Grammarly’s own tokenization explainer at https://www.grammarly.com/blog/ai/what-is-tokenization/.
The grammar-check layer is more complex. It parses the sentence structure — identifying subject, verb, object, clause relationships — then applies a set of rules (subject-verb agreement, comma splices, run-ons) and a statistical/neural model trained on pairs of incorrect and corrected sentences. This is grammatical error correction, or GEC: the model learned what errors look like by studying millions of examples of writers making them. For a clear GEC overview, see Trinka’s explainer at https://www.trinka.ai/blog/how-grammar-checkers-work/.
The style/tone layer is the newest addition, added in the LLM era, and the one most responsible for whole-sentence rewrites that erase student voice. It does not flag errors — it proposes alternatives it calculates as “clearer” by a learned metric. More on why that metric is not neutral in the section below.
This maps cleanly onto two AI4K12 Big Ideas. AI4K12 Big Idea #2 — Representation and Reasoning covers the rule-and-parse layer: the tool represents sentence structure and reasons from grammar rules. AI4K12 Big Idea #3 — Learning covers the ML layer: both the GEC model and the style model learned their patterns from data. Students are not just seeing grammar corrections; they are seeing two different AI architectures operating at once.
Rule-based vs. machine learning — what changed
Early grammar checkers were almost entirely rule-based. A programmer wrote explicit rules: if the sentence ends with a preposition, flag it; if the subject and verb do not agree in number, flag it. The rules could catch what they were written to catch and nothing else. They produced a lot of false positives — flagging correct sentences that broke a stylistic convention the programmer had treated as a rule.
The shift to machine learning, and then to neural transformer models, changed what the tool could do. A GEC model trained on large error datasets could learn patterns of incorrectness that no one had written a rule for. It could catch the subtler errors, generalize across sentence types, and — this is the part that matters for the classroom — suggest full rewrites, not just highlight problems. For a technical account of how GEC models evolved from rule-based to neural, see the BEA-2019 Shared Task paper at https://aclanthology.org/W19-4406.pdf.
This is exactly where the literacy gap widened. The tool got more capable at the same time that it got more opaque. When the spell-checker flagged “recieve,” the correction was transparent — the student could see the swap. When the style layer rewrites “The argument that Fitzgerald makes in chapter five is one that suggests” into “In chapter five, Fitzgerald argues,” the student might accept the rewrite without ever understanding why the original was weaker or whether the AI’s reading of what Fitzgerald “argues” is accurate. The tool’s growing fluency is the reason no one thought to hand you a user manual. That is not a failure of attention. It is a structural gap.
Where grammar checkers get it wrong

High precision, low recall, and a style layer with a point of view. These three failure modes look different in the classroom and require different teaching moves.
The precision and recall problem. A 2025 peer-reviewed study measuring GEC performance found approximately 91% precision (what the tool flags is usually a real error) but only about 51% recall (the tool misses roughly half of real errors), with an F1 score of approximately 0.66. Source: https://www.sciencedirect.com/science/article/pii/S1060374326000032. In classroom terms: the tool is reliable when it flags something, but students cannot assume that clean output means clean writing. A paragraph with no underlines may still have real errors — the model simply did not learn the pattern for them.
The false-positive / optional-flag problem. A peer-reviewed test of Grammarly on native-speaker scholarly writing found that 76% of “correctness” flags were optional — not actual errors — and 19% were outright false positives. Source: https://dept.writing.wisc.edu/blog/revisiting-grammarly/. Passive voice is the most common culprit: Grammarly flags passive constructions by default as if they are errors, when in academic and scientific writing they are often the appropriate choice. A student who accepts every passive-voice flag does not get a better argument — they get a more active-voice argument, which is not the same thing.
The dialect and style-layer problem. When the tool moves from grammar rules into style suggestions, it pushes writing toward Standard American English — and that push is not neutral. Research on NLP and LLM writing tools found that these systems flag African American English features at higher rates than equivalent features in Standard American English and silently rewrite inputs toward a single dialect baseline. Source: https://dl.acm.org/doi/fullHtml/10.1145/3531146.3533144. A student whose written voice draws on AAVE or another dialect will encounter more style flags than a student whose voice already conforms to Standard American English — not because one student writes worse, but because the model was trained on a particular kind of text and rewards resemblance to it.
Here is what these failure modes look like against real sentences:
| Sentence | What Grammarly flags | Is it an error? What a teacher could say |
|---|---|---|
| ”The book was written during a period of great change.” | Passive voice — suggests rewrite | Optional, not an error. Passive is appropriate here; ask: does the author matter more than the act? |
| ”He walked into the room, she followed.” | Comma splice | Genuine error — but a skilled writer could argue this is a stylistic choice (Hemingway does it). Context matters. |
| ”We was tired after the trip.” | Grammar error — subject-verb agreement | Genuine error in standard edited English; note this is also a feature of some dialects. |
| ”Her writing, it carries a kind of urgency.” | Style flag — suggests rewrite | Optional. “Topicalization” is a legitimate rhetorical feature. Accepting this flag erases a deliberate voice move. |
ISTE 1.3.b — evaluate the accuracy, perspective, credibility, and relevance of information — applies directly here. Students must evaluate the tool’s suggestions, not accept them. “Grammarly flagged it” is not a reason; “Grammarly flagged it and I agree because…” is the sentence that demonstrates writing growth.
Teaching students to use grammar checkers critically
The practical move is a three-option protocol on every underline: Accept, Revise, or Reject — and write one sentence explaining the choice. Not because documentation is the point, but because the explanation forces students to name the grammar concept at stake. A student who types “Rejected — passive voice is appropriate here because the subject of this sentence is irrelevant to the argument” has done more writing instruction in 15 words than any auto-accept.
A few classroom-ready moves:
- Print a paragraph with its Grammarly suggestions visible. Have students sort each suggestion into Accept / Revise / Reject in groups before anyone touches a keyboard.
- Give students a deliberately voice-rich paragraph — one with intentional fragments, a dialect feature, a comma-for-pause, an unconventional sentence opening. Run it through Grammarly together on the projector. Count the flags. Debate which are errors and which are choices.
- Draw the line between tool-assisted revision and tool-replaced revision: a student who used Grammarly to catch a comma splice and wrote the correction themselves revised their writing. A student who clicked “Accept all” and moved on did not.
For a full sequence that puts student voice at the center of AI writing feedback, see Teach Students to Revise AI Writing for Voice — Middle School and Evaluate AI Writing Feedback — Lesson for Middle School.
ISTE 1.1.c — use technology to seek feedback that informs and improves their practice — is the standard this protocol is hitting. Grammarly is a feedback tool. The instruction is in how students respond to the feedback, not in whether the tool is allowed.
CCSS.ELA-LITERACY.W.7.5 — develop and strengthen writing as needed by planning, revising, editing, rewriting, or trying a new approach — is the ELA anchor. Revising with Grammarly is an editing move only. It sits inside the larger revision process; it does not replace it.
For a printable lesson that walks students through the Accept / Revise / Reject protocol with a full argument-writing task, see the AI Peer Review and Evaluate AI Feedback lesson for grades 6-8 — a structured way to bring this critical-use framework into one period.
”Be the Grammar Checker” — a no-device activity

Before students use the tool critically, they need to know what the tool is doing. The fastest way to build that knowledge is to make them do it by hand.
Print one short paragraph — four to six sentences — with six to eight deliberately planted issues across the three layers: one misspelled word, two grammar errors (a comma splice, a subject-verb agreement), two style flags (a passive construction, a wordy phrase), and one legitimate dialect feature that a grammar checker would flag. Do not tell students which issues are which kind.
Minute-by-minute:
- 0-5 min (setup): Distribute the paragraph. Explain the three layers — spell-check (word-level dictionary), grammar (sentence structure rules), style (clarity and concision suggestions). Give each student three colored highlighters or pens.
- 5-20 min (sort): Students read the paragraph and mark each issue with the highlighter for its layer: one color per layer. They work in pairs. Each pair records their marks and a one-sentence justification for each flag.
- 20-30 min (debate): Whole-class share-out. For each flagged issue: is it a genuine error or a legitimate choice? Who would reject this flag and why? The planted dialect feature is the productive argument — expect disagreement, and let it run.
A typical class finishes the sort in 15 minutes and uses the remaining 10 for debate. The debate is the lesson. Students who argue that a flag should be rejected have to name the grammar concept — and that naming is more durable than any auto-correct.
This activity pairs with the printed AI Grammar Differentiated Worksheets for grades 6-8, which extends the three-layer analysis across eight grammar skills at three differentiation levels — a useful follow-on for the day after the unplugged activity.
If you want to extend the comparison further, the sibling mechanism post How Autocorrect Works — AI Lesson for Middle School covers the word-level prediction layer — a useful contrast to show students that the grammar checker is doing something structurally different (sentence-level parsing) from the phone’s autocorrect (next-word probability).
For the full voice-centered revision sequence, the AI Voice in Writing Revision lesson gives students a structured framework for identifying where Grammarly’s style suggestions erased something intentional — and how to get it back.
The gap was the literacy, not the shortcut
When a student’s paragraph comes back suddenly clean and neither of you can explain what changed, that is not a failure of vigilance. It is what happens when a three-layer machine runs on a piece of writing that no one has been shown how to interrogate. The gap was never the student reaching for a tool — it was that no one, student or teacher, was shown what the tool actually does. Maya, a former teacher who built these materials, designed the grammar worksheets and peer-review lessons because she kept running into the same wall: the AI was already in the room, and there was no curriculum for it.
Now there is. The grammar worksheets, the peer-review lesson, and the voice-revision unit are in the shop — printable, differentiated, and ready for the period after you hand back those suddenly clean paragraphs.
This post was drafted with AI assistance and human-finalized.
Quick questions
Grammarly runs three layers in sequence: a dictionary spell-checker, a grammar checker that parses sentence structure against rules plus a machine-learning model, and a style/tone layer that rewrites whole sentences. Students see the output as red, blue, and green underlines.
It has high precision — about 91% of what it flags is a real issue — but low recall, missing roughly half of real errors in a 2025 peer-reviewed study. A paragraph with no underlines can still contain errors the model never learned.
Yes, when they evaluate each suggestion instead of auto-accepting it. Teaching an Accept / Revise / Reject protocol — where students explain each choice — turns the tool from a shortcut into a thinking exercise.
Get the free AI-Proof Assignment Toolkit
10 ways to redesign any assignment so an AI chatbot structurally can't do it — plus a redesign worksheet, a 45-minute lesson, the “Spot AI Work” card, and parent templates. One email, all 5 pieces.
Prefer the full breakdown? See everything inside the toolkit →