How Grammarly Launders AI-Generated Content

Update 11/12/2025: I was contacted by a representative from Superhuman, the company that owns Grammarly. They said they are changing how “Humanized” text is displayed in the Authorship report. They are also working on ensuring that Gemini-generated text is correctly labeled. They hope for both changes to roll out next week.

Back in August 2024, Grammarly announced the launch of Grammarly Authorship, a new tool that would track how a work was written to prove it was not generated by AI. 

The goal was simple. By tracking what was typed, what was pasted, and how the document came together, it could show whether generative AI was likely used to create it and what portion is human-written.

However, the rollout of Grammarly Authorship has not been swift. Though I do not know when I received access to it on my account, I first noticed it in my Google Docs earlier this week. 

So, I decided to give it a shot. First, I used it on a legitimate piece of writing, namely my previous post, and then tested it with AI-generated content to see how it fared.

The results were a mixed bag. In some ways, Grammarly Authorship performed remarkably well; in others, it seemed confusing and could even make it seem as if it were using AI.

Still, I came away feeling as if this could be a powerful tool for those, in particular students, who fear being falsely accused of using AI. But this is a tool that requires proper use by both the author and the readers of the report. 

The First Test: A Real Post

Yesterday, I wrote an article about an expert witness who has sued several lawyers for ‘pirating’ her report. To test Grammarly Authorship, I enabled it as I started writing in Google Docs and continued writing as usual.

The report it generated contains many interesting details. It says that the post was 95% written by a human. It said 92% was typed directly and 3% was edited with spelling and grammar corrections. Of the 5% left, 2% was copied from a source, a CBS News article I cited, and 3% was “unknown”.

The unknown text was particularly interesting. It was the first sentence in the article. I brought that in from a different editor, but I had not turned on Grammarly Authorship when I copied it. However, I had turned it on before I pasted it. As such, it was “copied from an unknown source.”

Grammarly said I had spent 1 hour and 40 minutes on the post, which is roughly correct, and that there were no AI or plagiarism alerts. The report also includes an “Authoring Replay” feature that shows how it was written over time.

For the most part, the report is boring. It’s a mix of typed, edited and grammar/spelling corrected passages. Grammarly Authorship’s findings were accurate regarding what I did and how I wrote the piece. 

But then I turned my attention to pieces that I didn’t write, and that’s where things began to fall apart.

AI, Plagiarism and Text Laundering

After that, I decided to try less-ethical writing. First, I opened a new Google Doc and clicked the “Generate Document’ button provided by Google. This uses Google’s Gemini to generate a document for me. 

I gave it a prompt to generate a 500-word essay bout the history of plagiarism detection. After it was done, I checked the Grammarly Authorship report, and it identified it as 100% unknown. Since the text wasn’t pasted and was generated inside the document, it had no idea where it came from.

As bad as that was, it got worse when I moved it over to a new document and used Grammarly’s “Humanize” feature. This feature attempts to rewrite AI-generated text to sound more human. When I ran the report on that, it said that the document was “75% Typed by a Human.”

In the small print, it said that it was 0% typed by a human and 75% “Rephrased with Grammarly’s AI”. But the text was still flagged as green, and a casual observer could easily be fooled into thinking that it was mostly human-written.

For my final test, I gave ChatGPT a similar prompt and pasted the result into Google Docs. Grammarly Authorship immediately identified the text as AI-generated, noting that it came from ChatGPT. I then did something similar with an article from Plagiarism Today, and Grammarly Authorship immediately flagged the work as copied from a website. 

This raises two key problems with Grammarly Authorship. First, it’s easy to launder copied or AI-generated content through the system, and second, it requires those viewing reports to understand how they are created and what they should focus on when viewing them.

The Big Problems with Grammarly Authorship

Ever since the rise of generative AI, Grammarly has tried to walk a tightrope. On one side, it is a writing tool that aims to be a general writing assistant, incorporating AI to both improve and speed up the writing process. On the other hand, its popularity in education means that it doesn’t want to be seen as enabling AI cheating.

However, with Grammarly Authorship, it had to make a decision. How will it classify content rewritten/rephrased by its own AI? It chose to classify it as human-written. This means that, if you copy and paste content from another source, either AI or human-written, and run Grammarly’s “Humanize” tool over it, Grammarly Authorship will mark it as “Typed by a Human.”

To be clear, there are ways to see past this. First, the humanized text is labeled as “Rephrased with Grammarly’s AI” in the fine print. Second, the writing stats indicate that it took less than a minute or two to write the posts in question. This is, to put it mildly, physically impossible without cheating.

Finally, you can use the “Authoring Replay” feature to see how the work was created and quickly tell that it was copied from another source and rephrased.

But the problem is that those require effort and understanding from those reviewing the report. As we’ve seen with traditional plagiarism-detection software, instructors often blindly trust its findings, not digging deeper into the details of the report. This frequently leads to both false positives and false negatives.

Grammarly Authorship makes it easy for students to exploit this by simply passing their AI-generated content through Grammarly’s “Humanize” tool. It’s worrisome and gives the appearance of self-dealing.

However, it’s not the only limitation of the system.

Other Limitations

Beyond the issue with the “Humanize” tool, Grammarly Authorship has other limitations.

The biggest issue is that it can’t always track where the content is pasted from. Any material that doesn’t come from a website as a straight copy-and-paste isn’t fully tracked.

So, if you were to generate text in ChatGPT, copy it into a text editor, and then copy and paste it out of that, Grammarly Authorship would flag it as “unknown.” While this should still be considered suspicious (as all unknown text should), it raises fewer red flags than text that is clearly from another site or AI-generated.

This is because Grammarly Authorship in Google Docs relies on the Grammarly browser extension. As such, it can’t track anything from outside the browser. This is a technical limitation, but it still requires those reading the reports to be aware of the issue.

Second, when dealing with unknown text, it falls back onto more traditional AI detection that isn’t perfect. For example, in my piece generated via Gemini, many of the passages were highlighted as “unknown” but said, “It doesn’t match anything in our references or contain common AI text patterns.”

Still, it flagged the majority of the article as containing AI patterns, even though it listed it as unknown in the header.

Bottom Line

My concerns and issues aside, if I were a student doing legitimate work, I would enable this on every assignment. 

For all its problems, it did a decent job of verifying that my human-written work was not AI-generated. If questions arose about any of my work, this would be an easy counter to them.

If I were an instructor, I would learn how to read these reports and what actually to focus on. Though Grammarly Authorship gives neat-looking percentages, the writing stats and the authoring replay are far more helpful. They will tell you how long it took to write a piece and how it came together. 

Meanwhile, the percentages can be very misleading, especially if the student uses Grammarly’s “Humanize” tool. 

Unfortunately, that issue undermines Grammarly Authorship’s stated goal. It’s meant to be a tool to verify human authorship. However, if it can be easily subverted using Grammarly’s own tools, that creates mistrust.

Personally, I would move the “Rephrased with Grammarly’s AI” subhead to a new category. There is a wide gap between things that are “Typed by a human” and “Copied from a source.” A “Yellow” category indicating automated editing could significantly improve clarity.

All in all, the system has considerable promise. It’s just a shame that it decided to handle “humanized” text in a way that effectively launders it. 

This could be easily fixed later. But in its current form, those viewing these reports need to do so with great care and with an understanding how they were generated.