Do colleges count Grammarly as AI?

The Feature Line: Where Grammarly Stops Being 'Editing' and Starts Being 'AI'

If you're a student staring at your essay in Grammarly, wondering whether clicking "rewrite this sentence" just turned you into an AI cheat, you're asking the right question — and the answer is more precise than most policy documents make it sound. Colleges draw a feature line: standard grammar and spell-checking sits on the "editing" side, while Grammarly's generative rewriting tools sit squarely on the "AI" side.

This distinction isn't folklore passed around dorm rooms. It's written into official policies at some of the most selective institutions in the country. Yale's AI policy statement says that using a platform to review grammar or spelling "does not constitute application fraud." Brown's broad AI ban carves out an exception for basic spelling and grammar review, Cornell lists grammar and spelling review of completed essays as ethical use, and Caltech explicitly permits grammar and spelling checks — even while banning AI content generation entirely.

The Common App reinforces the same line at scale. It defines application fraud to include the "substantive content or output of an artificial intelligence platform," a rule that binds applicants to over 1,000 member colleges. Notice the wording: it targets an AI platform's output, not a red squiggle fixing "definately" into "definitely."

For coursework, the University of Illinois Chicago puts it more bluntly than anywhere else. According to UIC's official student guide, "Grammarly uses generative AI in any operation that rewrites text the student has already written" — and generative features require permission, summarized in the guide as "You need permission to prompt." Standard editing and proofing, by contrast, is treated as widely accepted fair use across UIC's colleges.

So where exactly is the line? Think of Grammarly's features in three tiers:

• Spell-check and basic grammar suggestions — safe at nearly every institution, and often explicitly named in policy exceptions.
• Clarity, conciseness, and tone suggestions — usually permitted editing, though some departments may restrict even these.
• Generative rewrites and GrammarlyGO prompts — treated as AI-generated content, requiring permission and disclosure wherever generative AI is restricted.

The stakes of guessing wrong are real. In 2024, a University of North Georgia student was placed on academic probation after her paper was flagged as AI-generated; she said she used only standard Grammarly features, and the case went viral. On Turnitin's own forum, faculty reported AI scores of 64%, 73%, and even 100% on student work claimed to be Grammarly-only — though informal faculty testing found grammar-only edits typically scored 0% or very low, suggesting the flagged cases likely involved generative rewrites.

The lesson: the feature you click determines the category you're in. Check your institution's specific policy before using anything beyond basic grammar suggestions, and disclose generative use early if it's permitted at all.

That same principle — know exactly what a tool does before you plug it in — is why SkillzBench is building a free directory that security-scans and benchmarks AI agent skills, so you always know what you're working with. Join the waitlist to get access when it launches.

When Grammarly Gets Students Flagged: Detectors, False Positives, and Real Cases

In 2024, a University of North Georgia student named Marley Stevens landed on academic probation after Turnitin flagged her paper as AI-generated — she insisted she had only used standard Grammarly features. The case went viral on TikTok, and the university temporarily pulled Grammarly from its recommended resources. It is the clearest warning of what happens when a permitted tool produces a forbidden pattern.

The core problem is that detectors do not see tools; they see writing. AI detectors like Turnitin analyze sentence uniformity, predictability, and perplexity — statistical smoothness — rather than which software touched the text. Heavy Grammarly rewriting can standardize syntax and replace natural human variation, so even fully original ideas can end up resembling AI output. As one analysis puts it, "the pattern itself can be flagged" — and there is no exact threshold for when writing "stops" sounding human. Institutional sensitivity settings also vary, so the same edits can trigger different AI scores from one class to the next.

The documented cases are sobering:

• On Turnitin's official forum, faculty reported AI scores of 64%, 73%, and 100% on student work claimed to be Grammarly-only. Turnitin acknowledged receiving "several similar reports" and said it was running tests.
• A 2025 study in the International Journal of Educational Technology in Higher Education found a human-written abstract edited with Grammarly received a 16% AI score.
• A SAGE Journals study that same year found detection accuracy across ten detectors ranged from 0% to 100% on AI text paraphrased through Grammarly Premium's "Improve it" feature — wildly inconsistent results.
• Forum skeptics noted that pre-AI-era personal writing was sometimes flagged as AI, with one user calling the detectors "useless."

There is an important counterpoint, though. Informal faculty testing found that text edited only with Grammarly's grammar suggestions scored 0% or very low on Turnitin — leading one professor to suspect that flagged students had actually used GrammarlyGo, the generative rewriting feature. Faculty on the forum drew the same line: "I am not going to fault a student for using Grammarly, but man there really is IMO a big difference between this and AI." The likely reconciliation across all these cases is that outcomes depend on which Grammarly features were used and how heavily.

The practical lesson is that a high AI score is a prompt to investigate, not proof of anything — and the best defense is provable authorship. Faculty workarounds documented on the forum include keeping pre-edited drafts as evidence and cross-checking with other detectors; AdvocatED recommends arguing authorship with records like Google Docs version history and Grammarly's Authorship report rather than arguing Grammarly "has nothing to do with AI." The same principle applies beyond campus: whether you are a student defending a paper or an agent builder choosing what to plug into your workflow, you need to know exactly what a tool does before you rely on it. That trust gap is what SkillzBench is building against — a free directory of agent skills where every version is pinned, security-scanned, and benchmarked with published lift scores, so nothing touches your work unverified. Skills, proven. Join the waitlist to be first in when it launches.

How to Use Grammarly Without Risking Your Grade or Application

Getting flagged for AI writing you actually wrote is a process problem, not a character problem — and process problems have process solutions. Here is the playbook that keeps you on the right side of the line.

First, read your specific institution's policy — and your course syllabus. The median school's AI admissions policy is just 32 words long, and most academic integrity policies are vague or undefined about where editing ends and generation begins. At UIC, for example, standard proofreading is "widely accepted," but rewriting functions require permission — and even departments within the same university disagree, with some language programs banning Grammarly entirely. Vague policy means the burden of interpretation falls on you, so ask before assuming.

Second, stay on standard editing features unless you have explicit permission for rewrites. Faculty testing on Turnitin's official forum found text edited only with grammar suggestions scored 0% or very low on AI detection — while faculty reported scores of 64%, 73%, and even 100% on work students claimed was Grammarly-only. The likely difference: which features were used and how heavily. Grammar and spell-check are safe; sentence rewrites put you in generative territory.

Third, disclose generative use early. If you do use GrammarlyGO or rewriting tools with permission, say so up front. As student advocacy guidance puts it: "Disclosure handled early is a footnote; disclosure discovered late looks like concealment." A proactive note to your instructor costs nothing; a detector score discovered later can trigger an investigation.

Fourth, keep process evidence that proves authorship:

• Google Docs version history, showing your drafts evolving over time
• Grammarly's Authorship report, which documents what you wrote versus what the tool changed
• Pre-edited drafts, outlines, and notes — the same evidence faculty on Turnitin's forum recommend keeping

If you are ever flagged, advocacy experts recommend arguing authorship with provable records rather than arguing that Grammarly "has nothing to do with AI." Remember, as one advocacy guide puts it, a high AI score is a prompt to investigate, not proof of anything.

The broader lesson applies beyond student writing: know exactly what a tool does before you plug it into your workflow. That same verification discipline drives SkillzBench, which is building a free directory where AI agent skills are security-scanned and benchmarked so users know what they are actually getting — the same transparency that would have prevented most Grammarly false-positive headaches.

The Bigger Lesson: Know Exactly What Your AI Tools Actually Do

The Grammarly question reveals something bigger: most of us use AI tools daily without knowing precisely what they do under the hood. UIC's guide draws the line at a specific technical boundary — "Grammarly uses generative AI in any operation that rewrites text the student has already written" — and that same line matters far beyond campus.

This "what does this tool actually do?" question now applies everywhere in an AI-assisted workflow. The research shows the stakes of not asking: a 2025 study found detection accuracy across ten AI detectors ranged from 0% to 100% on the same paraphrased text, and faculty on Turnitin's own forum reported AI scores of 64%, 73%, and even 100% on work students claimed was Grammarly-only. The tool's behavior — not your intent — determined the outcome.

The same blind spot exists in a newer corner of AI: agent skills. These SKILL.md instruction packages plug into coding agents like Claude Code, Codex, Cursor, and Gemini CLI, silently changing how your agent behaves. And the risk isn't theoretical — published research found roughly 26% of skills carry at least one vulnerability. Yet almost nobody verifies what they're installing before handing it access to their workflow.

That verification gap is exactly what SkillzBench is being built to close. Its free directory will security-scan, license-check, and quality-judge every skill, pin each version to a commit SHA, and benchmark each one in a sandbox — testing the same model with and without the skill to measure the actual Lift. A trust score from 0–100 combines security, quality, license, maintenance, and community signals, and money never touches the scores.

Before you plug anything into your stack, ask the same questions colleges ask about Grammarly:

• What does this tool actually change — suggestions, or full rewrites and actions?
• Where does the output go, and who else can see or flag it?
• Is it security-scanned, versioned, and independently benchmarked?
• Can you prove what you did versus what the tool did?

As AdvocatED puts it, "a high AI score is a prompt to investigate, not proof of anything" — but investigation only goes well when you have answers. Knowing exactly what your tools do is the new baseline for working with AI, whether the tool polishes your prose or powers your agent.

SkillzBench is launching soon — everything free: the directory, the MCP connector, benchmarks, and alerts. Join the waitlist at skillzbench.com and be first to browse verified, benchmarked agent skills.