How do I turn off AI on Grammarly?

Why You Can't Fully Turn Off Grammarly's AI — and What You Can Control

Here's the uncomfortable truth: you cannot turn off AI on Grammarly entirely, because AI isn't a feature bolted onto the product — it is the product. Grammarly has used artificial intelligence since its founding in 2009, and the machine learning that powers its spelling, grammar, clarity, and tone suggestions is inseparable from the tool itself. According to Grammarly's own support documentation, this foundational AI "cannot be turned off separately from Grammarly product offerings."

That said, not all of Grammarly's AI is locked in. The company distinguishes between two tiers:

• Foundational AI — the machine learning and natural language processing behind grammar, spelling, clarity, conciseness, and tone suggestions. This cannot be disabled, though Grammarly notes it "does not change the substance of the writing."
• Generative AI features — introduced in 2023, these produce new content from prompts and can meaningfully change what you write. These are off by default and can be toggled off entirely.

If your goal is to switch off the generative side, the steps are straightforward. As Leeds Beckett University's library guide explains: select "Account" from the menu, go to "Settings," open the "Feature Customisation" tab, and toggle the generative AI functions off — the check mark disappears once they're disabled. You can spot generative features by their pencil and lightbulb icons labelled "rewrite with Grammarly" and "Improve it."

Why does this control matter so much? Because Grammarly has spent the past few years rebranding itself as a full AI platform, adding AI paraphrasing, a humanizer, a GPT-based chatbot, reader reactions, and AI detectors. The company's CEO, Shishir Mehrotra, describes a mission to "solve the 'last mile of AI' by bringing AI directly to where people work." The more the platform expands, the more settings users need to actively manage.

The stakes became especially clear in March 2026, when Grammarly's parent company Superhuman disabled its "Expert Review" feature platform-wide after writers — including Stephen King — discovered their personas had been mimicked without consent, prompting a class-action lawsuit. As The Verge reported, Superhuman admitted "we clearly missed the mark." That one no longer needs disabling, but it's a reminder that AI features can arrive whether or not you asked for them.

One practical workaround: on the Grammarly website, choosing "classic" documents instead of "default" ones de-emphasizes the AI tools in the sidebar. And if you want AI assistance on your own terms — deciding exactly which tools an agent sees, rather than accepting whatever ships by default — that's the same principle behind SkillzBench, where you'll be able to browse, verify, and toggle free agent skills per agent. Control, not avoidance, is the realistic goal.

Step-by-Step: How to Disable Grammarly's Generative AI Features

Here's the good news buried in Grammarly's AI-everything redesign: the generative AI features can actually be switched off, and it takes less than two minutes. The catch is knowing where to look, because the setting isn't in the editor — it's tucked into your account settings.

Leeds Beckett University's library, which published clear instructions after students raised concerns, lays out the process in three steps. First, select "Account" from the menu. Second, go to "Settings" and then "Feature Customisation." Third, toggle the generative AI functions off — when a feature is disabled, the tick or check mark next to it disappears. The university notes that the exact steps may vary depending on your version or platform, so expect slight differences between the website, browser extension, and desktop app.

Before you toggle, it helps to know what you're looking for. Generative AI features are visually easy to spot once you know the signs:

• Pencil and lightbulb icons attached to prompts like "rewrite with Grammarly" and "Improve it"
• A dedicated "Generative AI" button on the Grammarly website
• Sidebar tools such as the AI humanizer, AI grader, and reader reactions — part of Grammarly's broader push into AI features that began integrating generative tools in August 2025

One feature you no longer need to hunt down is Expert Review. Grammarly disabled it platform-wide in March 2026 after writers — including Stephen King and the estate of Carl Sagan — objected to having their personas mimicked without consent, and after a class-action lawsuit was filed. So if you were looking for that button in the sidebar, it's already gone.

A word of caution for students and academic writers: Leeds Beckett requires students to acknowledge any use of Grammarly's generative AI in submitted work, even when it's used to support a disability or language need. Unacknowledged use "may be considered academic dishonesty." That's a strong argument for toggling these features off entirely if you're worried about accidentally leaning on them.

It's also worth remembering what this switch does and doesn't control. Grammarly's foundational AI — the machine learning behind spelling, grammar, clarity, and tone suggestions — cannot be turned off separately from the product itself. Per Grammarly's own support documentation, that layer "does not change the substance of the writing," while generative AI can. The Feature Customisation toggle only governs the second, content-generating tier.

For readers who want this kind of control more broadly — deciding exactly which AI capabilities an assistant can see and use — that's the same principle behind SkillzBench, which is building a free directory of verified agent skills with per-agent toggles. You can join the waitlist to be notified at launch. In the meantime, those three clicks in Feature Customisation will give you a Grammarly that corrects your writing without ever trying to write it for you.

Reducing Grammarly's AI Footprint: Classic Documents and the Expert Review Saga

Even after you've toggled off Grammarly's generative AI features, the platform still nudges AI toward the center of your writing workflow. Fortunately, there are a few workarounds that shrink AI's footprint further — and one controversial feature you no longer need to worry about at all.

On the Grammarly website, new documents come in two flavors. "Default" documents foreground AI tools throughout the interface, while "classic" documents use a simpler three-tab sidebar covering suggestions, generative AI writing, and AI/plagiarism detectors. A hands-on academic review from Carleton College found that choosing classic documents de-emphasizes the AI tooling, keeping the focus on traditional grammar and clarity suggestions instead.

There is a trade-off, though. The same review notes that Grammarly's browser extension and desktop app require access to a hefty amount of personal data. If minimizing Grammarly's reach is your priority, working in classic documents directly on the website is the leaner option.

Expert Review, the AI feature that generated edit suggestions "inspired by" real writers — including Stephen King and the late Carl Sagan — without their permission, was disabled platform-wide in March 2026. Grammarly's parent company, Superhuman, initially offered an email opt-out for affected writers, then acknowledged it "didn't go far enough" and pulled the feature entirely while it "reimagines" it.

The removal followed a class-action lawsuit filed in the Southern District of New York, with damages alleged to exceed $5 million (£3.7 million) over the misappropriation of "hundreds" of writers' identities. Per BBC reporting, more than 40 people contacted the lawyers within 24 hours of the suit being filed. Lead plaintiff Julia Angwin called the AI-generated edits a "slopperganger," noting the edits "were not good... they were making the sentences worse."

The practical takeaway: if you were hunting through settings for an Expert Review toggle, you can stop looking. Key facts to remember:

• Expert Review was removed for all users in March 2026 — no individual action needed.
• Superhuman's director of product management admitted the company "clearly missed the mark."
• The CEO claimed the feature "had very little usage," though critics tied the removal to backlash and litigation.

The Expert Review episode highlights why many users want granular control over which AI capabilities touch their work. If that resonates, the same principle applies to AI agents more broadly: choosing exactly which skills and tools an agent can see matters. SkillzBench is building a free directory and MCP connector that lets you do precisely that — browse, verify, and toggle agent skills on your own terms.

AI Detectors, Academic Integrity, and When Disabling AI Really Matters

For many readers, switching off Grammarly's AI isn't about preference — it's about protection. The stakes are highest where AI detectors and academic integrity rules come into play, and the picture there is messier than most students realize.

Start with the detectors themselves. Grammarly's own support documentation states that no AI detector today can reliably and conclusively determine whether AI was used to produce text, and that detector results should be treated as a single data point combined with manual review. In other words, the very tools institutions lean on are, by the vendor's own admission, inconclusive.

The consequences fall unevenly. Research cited in a Carleton College review indicates that AI detectors are more likely to produce false positives for students who write in English as a second language — meaning the students most likely to use grammar support are also the most likely to be wrongly flagged. That's a strong argument for disabling generative AI features entirely rather than trying to prove after the fact that you didn't use them.

Academic integrity rules add another layer. Leeds Beckett University, for example, requires students to acknowledge any use of Grammarly's generative AI features in academic work — even when those features support a disability or language need — and warns that failing to acknowledge such use may be considered academic dishonesty. If accidental AI use could put your grade at risk, toggling the features off via Account → Settings → Feature Customisation is the safest move.

For those who need attribution rather than removal, Grammarly recommends its Authorship feature, which tracks your writing process and documents proper attribution — useful evidence if a detector ever questions your work.

Before you decide, consider what you're actually turning off:

• Foundational AI (grammar, spelling, tone) cannot be disabled — it's integral to the product.
• Generative AI features can be toggled off in Feature Customisation settings.
• Classic documents on the Grammarly website de-emphasize AI tools compared to default documents.
• Expert Review was disabled platform-wide in March 2026, so that one no longer requires action.

The broader lesson is about control. Grammarly markets its generative AI as giving users "full control," but that control applies only to the generative layer — and knowing exactly which AI tools touch your work matters more as institutions tighten their rules. The same principle applies beyond writing: when you want AI assistance on your own terms, choosing precisely which skills and tools an agent can see is the way forward. SkillzBench is building exactly that — a free directory of vetted agent skills with per-agent toggles, plus an MCP connector so you decide what your AI sees. The waitlist is open now if you want AI you can actually govern.

Take Control: Curating AI Assistance on Your Own Terms

The Expert Review saga is the clearest lesson yet in why AI assistance should never arrive uninvited. Grammarly's parent company disabled the feature platform-wide in March 2026 only after a class-action lawsuit alleging damages exceeding $5m and backlash from writers whose personas were mimicked without consent (BBC News). The lesson for professionals: don't wait for a vendor's off-switch — curate your AI from the start.

Grammarly's own support documentation frames its AI as giving users "full control," yet the foundational AI powering grammar and tone suggestions cannot be turned off at all (per Grammarly). Only the generative layer introduced in 2023 is user-controllable via Feature Customisation (Leeds Beckett University). That gap between marketing language and actual capability is exactly what agent builders and business users should watch for in every AI tool they adopt.

The stakes extend beyond annoyance. Leeds Beckett warns students that unacknowledged use of Grammarly's generative features "may be considered academic dishonesty" — even when used to support a disability or language need. And with independent research finding that roughly 26% of published agent skills carry at least one vulnerability, blindly connecting agents to whatever tools are available carries real risk.

This is the philosophy behind SkillzBench, AIQ Labs' free, independent directory of AI agent skills. Instead of accepting whatever AI a vendor bundles in, SkillzBench's approach inverts the model:

• Discover and verify — skills are crawled from GitHub, deduplicated, pinned to commit SHAs, license-checked, security-scanned, and quality-judged before anything else happens.
• Benchmark in a sandbox — every skill is tested with the same model, with and without the skill, so the Lift it actually provides is published in percentage points.
• Equip per agent — a free MCP connector lets you toggle skills and tools for each agent individually, so the agent only ever sees what you deliberately chose.

That last point is the Grammarly lesson, applied. When the Expert Review controversy broke, writers had no per-feature control — the feature was simply there until public pressure removed it. A loadout model, where each agent sees only the skills you selected, makes unwanted AI behavior structurally impossible rather than something you discover after the fact.

SkillzBench is pre-launch, with a live waitlist and newsletter for early access. Everything on the platform — the directory, the MCP connector, benchmarks, and alerts — will be free, with trust scores kept separate from any commercial influence. For professionals who read this article because an AI feature arrived unannounced, that's the alternative worth watching: assistance on your terms, not the vendor's.