Grammarly Revisited

29 October 2025
By Kieran Barker '25

Update 2/17/26: Since this post was written in October 2025, Grammarly has added a couple of new features. Below is a video walkthrough of the updated suite of features. Aside from the addition of this video, the post below has not been changed.

Next to names like ChatGPT, Copilot, or Claude, Grammarly might seem like an odd inclusion. The former, after all, are some of the big names of today’s GenAI craze. The latter? Ostensibly, a writing tool of yesteryear. In fact, at the time of writing, Grammarly is one of the most-used GenAI tools, edging out both Copilot and Claude (though not ChatGPT) in web traffic volume.

For those who have missed it, Grammarly has spent the past few years rebranding itself as an AI platform. And while it has maintained its identity as a writing assistant, former users may be surprised at the extent to which this move has transformed Grammarly’s service. On top of the proofreading tool that made it famous, Grammarly now includes AI paraphrasing, AI “grading,” AI “expert review,” and even a built-in GPT-based chatbot, among several other AI features. 

In this post, I’ll break down these changes and give a few things to consider about the new Grammarly. I’ll start with an overview of Grammarly’s basic functionality, then move on to a closer look at its new AI features. In the last section, I’ll give my overall thoughts about these changes.

The Basics

Grammarly’s tools can be accessed on its website, on web-based word processors (like Google Docs) through a browser extension, or on any desktop application (like Word or Notes) through the Grammarly app. To keep this post to the point—but also because the browser extension and desktop app require access to a hefty amount of personal data—I will be focusing on the tools as presented on the website. 

Grammarly’s website is set up a bit like Google Docs or Word. From the home page, you can create or upload new documents, of which there are two kinds: default and classic. Default documents foreground Grammarly’s new AI tools more than their classic counterparts, but both work in more or less the same way. Documents are divided into a space for adding text and a side bar with Grammarly’s tools. The text section has limited word processing features, making it clear that these documents are less for drafting text than for using Grammarly’s tools. The side bar varies slightly between document types. On default documents, it exclusively houses the new AI writing tools. On classic documents, it’s broken into three tabs: one for reviewing suggestions, one for writing with GenAI, and another for Grammarly’s AI and plagiarism detectors.

Once you have written or pasted in something, Grammarly will mark up the document with suggestions, just like Docs or Word. These are Grammarly’s general, mechanical suggestions about conciseness, spelling, and grammar. You can act on them by clicking the underlined text. If you want to use one of the AI features, you can click on the sidebar and, after filling in any necessary fields, it will start giving you revision suggestions. These suggestions pop up one at a time, paragraph by paragraph, asking you to approve or discard the suggestions as you go. Except for the AI chatbot, all of the AI features share this basic structure.

With this background in place, I want to take a closer look at these individual additions. Since there are quite a few new AI features, I won’t spend too much time on any one. Instead, I’ll give an overview of how each feature works, along with some observations on their usefulness.

New Features

AI Chat

“AI Chat” is a built-in GPT-based GenAI chatbot. This means that it works like a less powerful version of ChatGPT, Gemini, or any other chatbot, except that it can also read from and edit the text section of your document. You can ask it to make edits or to write new text, either from scratch or on top of what you have written. You can also ask it factual questions, like “who was Friedrich Schiller?” Being a scaled-down model, though, Its responses to questions like these are noticeably abridged compared to responses from ChatGPT or Gemini. 

Grammarly Proofreader

The proofreader is basically a souped-up version of Grammarly’s ordinary suggestions. After you have entered in your text, it will display a handful of suggestion categories, like “shorten sentences,” “tighten word choice,” “smooth transitions,” or “clarify key argument.” Selecting a category will initiate the paragraph-by-paragraph style suggestions, and Grammarly will go through your document making suggestions of the kind selected.

The suggestion categories offered are not the same each time; Grammarly displays them based on details about your text. The exact mechanism behind this is obscure, but one thing it seems to look for is the kind of writing at hand. When I gave it a passage from a work of fiction, for instance, the categories changed to “clarify physical contrasts” and “enhance narrative tone.” Incidentally, the suggestions it gave for literary prose were much poorer than those it gave for formal, more mechanical writing. This shouldn’t be too surprising, since formal writing is typically more homogenous and patterned than literary prose, making it more easily imitable by pattern-seeking GenAI.

Paraphraser & Humanizer

The paraphraser and humanizer tools do essentially the same thing: rewrite text in a different tone. The only real difference appears to be that the humanizer is more specifically intended to make AI-generated text sound more natural. Both work the same way, however, and basically make a redundant pair (the paraphraser even has an option called “humanize”).

When you select either one, you can choose a “voice” in which it will rewrite your paper. Grammarly provides a few sample voices (“streamlined,” “creative,” “the scholar,” etc.), but you can also create your own voice preset by submitting your own example texts. Grammarly then uses these texts as a template on which to model its suggestions. Once again, these suggestions can be helpful for certain kinds of formal or rote writing, but they are much weaker for kinds of writing where more creativity, variation, or detail is necessary.

Expert Review

The “expert review” tool is the first of Grammarly’s truly bizarre features. It works by going through your paper and giving revision suggestions “inspired” by the ideas of experts it classifies as relevant to your topic. The name attached to them notwithstanding, these suggestions are usually quite basic. After reading an old philosophy paper of mine on a Platonic dialogue, for example, expert review suggested that I illustrate Socrates’s argument with a modern example, an idea it claimed was inspired by John Dewey. Not bad advice per se, but also not expert-level advice. Moreover, these suggestions are not context-sensitive. Grammarly does not generate them with the specific aim of your writing in mind; it can only give generic suggestions based on the subject matter of your writing.

Reader Reactions

The “reader reactions” tool makes predictions of what readers of your document will take away and what they will have questions about. In lieu of an actual person to read your text, I could see this feature being helpful as a way to get out of your author’s bubble. Like all GenAI output, though, these are only predictions based on how the LLM breaks down your text. While potentially useful for stimulating thought, they can’t replace the unique insight and context-sensitivity of a human reader.

Like the expert review tool, this one can also add text that it claims addresses these questions. In my experience, though, this text almost always felt out of place in the surrounding writing, not unlike other kinds of AI-generated writing.

Citation Finder

The citation finder searches your document for claims in need of citation. It breaks down claims into three types: supported, debated, and contested. The inclusion of the latter two categories also makes it a kind of rudimentary fact-checker. Once it finds a claim in need of citation (or correction), it directs you to a source on the internet that you can use to cite the claim. The quality of these sources varies. It sometimes suggests sources like Forbes or Wikipedia, and other times suggests encyclopedias or arxiv articles. If you select the suggested source, Grammarly can write you a citation in MLA, APA, or Chicago format.

AI Detector & Plagiarism Checker (exclusive to Pro)

The AI detector and plagiarism checker are the only two AI features wholly exclusive to Grammarly Pro, Grammarly’s paid version. The AI detector scans your document and shows you passages that it says could be flagged as AI-generated text. It also offers to write a citation for those passages.

The plagiarism checker does more or less the same thing, except that the citations it offers are for websites with content similar to your paper’s. The filter for these sources is much less discerning than the citation finder, though, and it will often suggest articles from strange or otherwise unreviewed websites.

With both, it is difficult to say how accurate the detection is. In my experience testing the AI detector, I found that it could somewhat reliably detect AI-generated or edited prose (including text “humanized” by Grammarly!). It struggled to detect other kinds of AI-generated text like bulleted lists or brief notes, however. The plagiarism detector was equally inconsistent, and I caught it giving me both false positives and negatives. Note that Grammarly itself cautions against solely relying upon AI and plagiarism detectors to evaluate text. The need for this warning is further bolstered by findings that suggest AI detectors are more likely to give false positives to writing from students for whom English is a second language.

AI Grader

The AI grader is probably the strangest of Grammarly’s AI tools. It is built to predict what grade your document would receive if it were graded for a class. To receive a “grade,” you can upload your assignment rubric, enter your class section, and enter your instructor’s name. Grammarly then searches the web for publicly available information about your instructor and that class (RateMyProfessor reviews, public syllabi, public assignments, etc.). Note that some of this information may have been uploaded without the consent of the instructor. It then tries to simulate the grade and feedback that professor would give. 

This feature performs very poorly on most fronts. For one, probably unsurprisingly, its simulated feedback is unrealistic. More often than not, Grammarly latches on to the instructor’s area of expertise and checks how much your writing engages with that topic. If your English professor specializes in Russian literature, for example, the AI grader might mark down your Shakespeare paper for not drawing connections to Tolstoy or Turgenev (this isn’t an exaggeration). Moreover, following these suggestions might result in your actual paper being marked down. The AI grader also gives you the option to just generate text that “fixes” these “problems,” another move that, if disallowed by your instructor’s syllabus, could get your paper marked down, or worse.

It is conceivable that someone could use the AI grader’s feedback in a similar way to the “reader reactions” from before. If it makes a point that you agree could improve your paper, it might be worth taking the suggestion on board. That being said, it strikes me that good suggestions are a little too infrequent to justify consistent use of the tool in this way.  

Takeaways

Although Grammarly markets itself as a writing assistant, many of these tools are designed to do your work for you. The fundamental architecture shared by all of Grammarly’s AI tools—the paragraph-by-paragraph suggestion model described above—makes this clear. This is unfortunate on at least two levels. For one, Grammarly’s tools are just generally more useful when used for stimulating thought, for giving another perspective, even if it isn’t that of a real human reader. These tools can still be used in this way, but so much of the focus and effort on Grammarly’s end seems to be on its ability to actually rewrite your text. For another, it obscures our sense of what kind of “assistance” is appropriate. As an antidote, it might be helpful to ask: “would it be acceptable for another person to do the work I’m about to have AI do?”

Another issue worth mentioning is the poor usability of Grammarly’s free version. If the plagiarism and AI detectors were the only features exclusive to Pro—or even if more individual features were restricted to Pro—Grammarly Free might be worth using for a few features. As it stands, though, over half of all suggestions are locked behind an advertisement for Grammarly Pro. Having done most of my initial testing on a free trial of Grammarly Pro, it was astonishing to see how cumbersome the free version was to use. At the same time, though, I can’t recommend upgrading to Grammarly Pro. In my experience with the platform, the tools on offer simply aren’t worth the price of a subscription.

The hit-or-miss usefulness of Grammarly’s tools tells an interesting story about AI and writing, though. I want to offer two tentative conclusions from the above observations: 1) AI tools seem to be at their best when dealing with more patterned, homogenous kinds of writing; and 2) specialized tools made for already skilled hands tend to fare better than generalized tools that focus on automaticity. We saw (1) at play with the proofreader and the paraphraser tools. We saw (2) on display in all the tools that can stimulate further thought in the user but struggle with generating quality corrections on their own. As the AI world continues to shift, these conclusions are of course subject to revision. It might turn out that an as yet unreleased tool will upend these observations. For the time being, though, it might be worth keeping observations like these in mind as we continue to experiment with AI.