Can teachers tell if you use Grammarly AI?

The False Flag Problem: When Grammarly Gets Innocent Students Accused

Picture this: you write an essay yourself, run it through Grammarly for a quick grammar check, and days later you're sitting in a dean's office accused of AI cheating. It sounds absurd, but it's happening — and the consequences can be life-changing.

Consider Marley Stevens, a University of North Georgia junior who received a zero on her criminal justice paper after Turnitin flagged it as AI-written. She says she only used Grammarly's standard grammar and spell-check extension. The stakes were real: she lost her HOPE Scholarship eligibility, and her story spread to over 5.5 million views on TikTok before Grammarly stepped in with a $4,000 contribution to her fundraiser.

Stevens isn't alone. A University of Lethbridge research group ran a known human-written abstract — one edited only with Grammarly's phrase-level assistance — through Turnitin and got a 16% AI-generated score. And at Shaker Heights High School in Ohio, junior Zi Shi's assignment was flagged by GPTZero, with his teacher suggesting Grammarly use triggered the flag.

The scale of the problem is growing. More than 40% of surveyed 6th–12th grade teachers used AI detection tools last school year, according to a Center for Democracy and Technology poll reported by NPR. That means millions of students are now submitting work to software that even the vendors admit shouldn't be treated as proof.

So why does a grammar checker trip AI alarms? The mechanism appears to be about how deeply you engage with suggestions:

• Blindly accepting major suggestions — especially sentence restructuring — makes text look machine-polished, which detectors read as AI writing.
• "Tinkering" with suggestions and rewriting them in your own words tends not to trigger flags, according to English professor Steven Marsden.
• The results are inconsistent even between tools: one Grammarly-assisted article scored 24.88% AI on ZeroGPT but 0% on Grammarly's own detector.
• Stevens' paper was flagged by Copyleaks at the time, yet when re-run later, the same tool concluded "This is human text."

The deeper lesson cuts both ways: detectors produce false flags on innocent work, while paraphrasing tools that enable actual misconduct often slip through undetected. That's a verification problem — trusting a tool's output without understanding what it actually measures. It's the same principle driving SkillzBench, which is building free security scanning and benchmarking for AI agent skills so users can see what a tool really does before relying on it.

Until detection improves, the burden falls on students to document their writing process — and on everyone to treat any single tool's verdict with healthy skepticism.

Why Detection Fails: Teachers and Tools Both Get It Wrong

Your teacher's gut instinct is not the lie detector you think it is — and neither is the software your school pays for. When researchers actually tested this, human judgment and automated detection both failed, in different directions, at scale.

In two randomized-controlled experiments, peer-reviewed research found that neither 89 novice teachers nor 200 experienced teachers could identify ChatGPT-generated essays among student work. Worse, they were overconfident while doing it — a combination that turns honest mistakes into confident accusations.

The software doesn't rescue the situation. Detector results are wildly inconsistent from tool to tool. One student journalist's article, written with Grammarly's help, scored 24.88% AI on ZeroGPT but 0% on Grammarly's own detector — the same text, opposite verdicts. Marley Stevens' flagged paper came back "This is human text" when she later re-ran it through Copyleaks, after she had already received a zero and lost her HOPE Scholarship eligibility.

The false-positive math is where this gets institutional. Grammarly's Jenny Maxwell laid it out plainly: even at 98% accuracy, a university processing 50,000 papers a year would still see roughly 1,000 falsely flagged papers annually. And Turnitin's own stated false positive rate is around 1 percent, while about 15 percent of bot-written text slips through unflagged — meaning the tools miss real misconduct while penalizing honest students.

The documented failure patterns include:

• Turnitin scored a known human-written, Grammarly-edited abstract at 16% AI, per University of Lethbridge research.
• An English teacher's own Ph.D. dissertation chapter scored 89–91% AI-written on GPTZero.
• A 30.76% probability score triggered a false accusation — below even Turnitin's own reliability guidance.

The paradox is that paraphrasing tools built to disguise AI use consistently evade detection, while Grammarly's readability enhancements trigger flags — especially for non-native speakers. Academic integrity researcher Mike Perkins puts it bluntly: these tools are "not fit for purpose."

The broader lesson applies far beyond plagiarism software: verify what your AI tools actually do before trusting their outputs. That's the same principle SkillzBench is building around for agent skills — security scanning, benchmarking, and a free MCP connector that lets you test claims instead of taking them on faith. Skills, proven — because blind trust in tools, whether detectors or agents, is exactly what produces these failures.

What Actually Tips Teachers Off: The Accept-vs-Tinker Line

If there's one practical rule that separates flagged papers from unflagged ones, it comes down to a single click. Teachers and detectors aren't looking for whether you used Grammarly — they're looking at how you responded to its suggestions.

English professor Steven Marsden puts it plainly: "If you tinker with something Grammarly suggests, then Turnitin won't mark it, but if you just click accept on big stuff like sentence structures… well, that is technically AI writing your stuff." Blindly accepting major suggestions — especially whole-sentence restructuring — is what reads as machine-generated. Editing a suggestion in your own words does not. Student reporting at Stephen F. Austin backs this up: Turnitin can apparently distinguish minor grammar tweaks from larger structural rewrites, and it's the structural rewrites that draw flags.

The irony is sharp. The University of Lethbridge research group found that paraphrasing tools — arguably more ethically fraught, since they rewrite text wholesale — "score low values in AI detection software, allowing the changes to go mostly unnoticed." Meanwhile, tools that "subtly enhance readability" like Grammarly trigger detection and false positives. The group hardest hit? Non-native English speakers, who rely on readability tools most and whose polished prose pattern-matches to AI output.

There's also an institutional double standard worth naming. Many schools ban ChatGPT outright while English departments actively recommend Grammarly — even though both are AI-powered. The distinction, when one is offered, is that ChatGPT can replace the learning while Grammarly refines it. At SFA, faculty currently "let students off the hook" if a student claims they only used Grammarly — a loophole that could just as easily enable cheating as protect honest writers.

What actually tips teachers off, then, looks like this:

• Wholesale acceptance of sentence-restructuring suggestions, which produces the uniform, machine-smooth prose detectors are trained to spot
• A detector score that crosses a teacher's personal threshold — one educator investigates anything above 50% likelihood, and finds roughly 75% of flagged students admit AI use
• A mismatch between the paper's polish and what a student can defend orally — flagged students at Lethbridge "easily passed informal tests" when asked to explain their own writing
• Revision history that shows a single burst of edits rather than genuine drafting over time

The broader lesson extends well beyond the classroom: know exactly what your AI tools are doing before you lean on them. That same verification-first mindset is what SkillzBench is building for AI agent skills — security scanning, sandboxed benchmarking, and a free MCP connector, all under development pre-launch. Trusting a tool blindly, whether it's a grammar checker or an agent skill, is where the trouble starts.

How to Use Grammarly Without Getting Flagged: A Student's Action Plan

Getting flagged is not the same as being guilty — and the research shows the difference often comes down to a few habits you can control. Here is a practical action plan, grounded in what educators and detector vendors themselves say works.

Step 1: Rewrite, don't accept. English professor Steven Marsden puts the mechanism plainly: tinkering with Grammarly's suggestions doesn't trigger Turnitin, but clicking accept on big sentence restructures is "technically AI writing your stuff." Treat every major suggestion as a prompt, not a command — retype the fix in your own phrasing, and keep your own sentence architecture intact.

Step 2: Document your process. Both academic integrity expert Tricia Bertram Gallant and Turnitin's chief product officer recommend students keep records of their tool use and the evolution of their work. That matters because teachers like John Grady check revision history — time spent, number of edits — before ever confronting a student, and only investigate further past a 50% detector threshold.

Step 3: Be ready to defend your work out loud. At the University of Lethbridge, where papers scoring at or above 30% AI triggered review, flagged students easily passed informal oral defenses of their own writing. If you can explain your argument, sources, and choices on the spot, a detector score loses most of its power.

Your quick pre-submission checklist:

• Rewrite major Grammarly suggestions in your own words instead of accepting them.
• Keep drafts, revision history, and a note of which tools you used and how.
• Run your final draft through multiple detectors — one student journalist's Grammarly-edited article scored 24.88% on ZeroGPT but 0% on Grammarly's own detector.
• Rehearse a two-minute verbal walkthrough of your thesis and evidence.

Know what a score actually means. Turnitin itself says scores of 20% or lower are less reliable, and that detection should be "a starting point for a conversation with a student, not a final ruling." Even at a stated ~1% false positive rate, Grammarly's Jenny Maxwell calculates a university processing 50,000 papers a year would still see roughly 1,000 falsely flagged submissions.

The broader lesson applies to any AI tool you rely on: verify what it actually does before you trust it. That's the same principle SkillzBench is applying to agent skills — security scanning, sandboxed benchmarking, and a free MCP connector, all in development ahead of launch — so the tools working on your behalf never work against you.

The Bigger Lesson: Verify What Your AI Tools Actually Do

The Grammarly story ends where a bigger principle begins: trust in AI tools has to be earned, not assumed. Blind faith fails in both directions — detectors flag honest students, while the same blind trust lets actual misconduct slip through undetected.

Consider the false positives first. Marley Stevens lost her HOPE Scholarship eligibility after Turnitin flagged a paper she says she only checked with Grammarly's standard extension. A University of Lethbridge study ran a known human-written, Grammarly-edited abstract through Turnitin and got a 16% AI score. Even Grammarly's own team points out the arithmetic: at 98% accuracy, a university grading 50,000 papers a year would falsely flag roughly 1,000 of them.

The false negatives are just as corrosive. The same Lethbridge research found that paraphrasing tools score low in AI detectors, letting their changes "go mostly unnoticed" — while readability tools like Grammarly trigger flags and increase false positives, especially for non-native speakers. Meanwhile, some faculty "let students off the hook" if they claim they used Grammarly, a loophole student journalists note could enable cheating. One English teacher's own Ph.D. dissertation chapter scored 89–91% AI-written on GPTZero — proof that a score alone proves nothing in either direction.

The lesson generalizes far beyond the classroom. Whether you're a student, an educator, or a business owner deploying AI at work, the failure mode is identical: relying on a tool whose behavior you haven't verified. Before you depend on any AI tool, ask:

• What does it actually do, and where's the line between assistance and generation?
• How reliable is it — and what's the documented error rate in both directions?
• Who benefits from trusting it blindly, and who pays when it's wrong?
• Can its claims be independently tested, not just taken on faith?

That verification mindset is exactly what SkillzBench is building for the next wave of AI tooling: agent skills. Published research found roughly 26% of agent skills carry at least one vulnerability, yet most people install them sight unseen. SkillzBench's pre-launch platform will security-scan every skill, pin each version to a commit SHA, benchmark each one in a sandbox with and without the skill, and publish the Lift separately from ratings — so you can see what a tool actually does before you rely on it. A free MCP connector will let you connect any agent to only the skills you've chosen and verified.

Detectors shouldn't be the final word on a student's work, and a skill's self-description shouldn't be the final word on its safety. Verify first, then trust. Skills, proven. Browse the free skills directory at SkillzBench and join the waitlist to be first in line when the platform launches.