Thanks to its AI system, Grammarly is now one of the most valuable US startups

This year has been big, momentous for Grammarly, the AI–based grammar and writing assistant startup.

In January, its ads went viral after a user posted a tongue-in-cheek dance, inspiring tens of thousands of TikToks. And now, as the year winds down, the company raised $200 million—at a $13 billion valuation, making it the No. 10 most valuable startup in the nation.

Underneath the company’s soaring valuation is a layered system of AI and natural language processing (NLP)—specifically tools like large language models, machine learning (ML), rules engines, and 12 years of user feedback—that it uses to make suggestions to its reported 30+ million users.

How it works

When Grammarly started out in 2009, it focused more on correcting the most common grammatical errors. Now, it offers edits on everything from tone and inclusive-language suggestions to conciseness and clarity.

• After a user writes an email, for example, the text is run through a “neural machine translation” process, according to the company, to generate a first take on potential changes.
• Then everything is fed into large language models, which use those initial results to try to draft the edited version of the text.
• After that, rules engines take a crack at the sentence—think: natural language processing rules about how the language works—before the final suggestions are shown to the user.

“We can say, ‘Okay, the output that the ML model created—does that actually make sense, or does it need to be tweaked a bit?’” Rahul Roy-Chowdhury, Grammarly’s global head of product, told us. “ML doesn’t always get it right. ML models are notoriously, sometimes, really bad at getting output right. So we layer on our knowledge of language models and NLP to fix it.”

Instead of combining the whole English language into one especially large language model, each aspect of language that Grammarly edits for—for example, tone, empathy, inclusive language, clarity, conciseness, and correctness—has its own machine learning “ecosystem,” Roy-Chowdhury explained.

Fresh funds

Grammarly says it will invest some of the $200 million it just raised in advancing user trust and personalization.

Since we speak differently on email than Slack, for example, Grammarly has plans to analyze the way its users typically communicate on different platforms and apply that to its suggestions moving forward. The company declined to comment on if users can opt out of this feature.

• “The more Grammarly can understand your world and your context...the more we can help you,” Roy-Chowdhury said.

The company is already compliant with standards from HIPAA and the California Consumer Privacy Act, Roy-Chowdhury said, but the team wants to continue making sure users have control over how their data is being used.

• Grammarly does not currently allow users to opt out of having some of their feedback, like whether they accept or reject suggestions, included in training data.

Looking ahead: “We just have a huge market opportunity ahead of us,” Roy-Chowdhury said. “Communication is complicated; we’ve barely scratched the surface of how we can help people communicate more effectively.”

Update/Editor's note: This story was updated on November 19, 2021, to clarify a quote from Rahul Roy-Chowdhury, Grammarly’s global head of product.