People using writing assistants at work might love sending polished emails to colleagues or crafting smarter company social media posts. Their company’s IT and legal teams might not love the fact that some of those tools use their content to train their AI.
When people use Grammarly, a popular writing software, they grant the company permission to use the content they write in the tool to help adjust and improve its machine-learning models. Another writing assistant, Bramework, has terms and conditions requiring users to give the company irrevocable, perpetual and worldwide rights to use, reproduce and distribute the content they write when using the system.
Grammarly uses human linguists along with AI tools such as natural-language processing and machine learning to automate suggestions for ways to make writing crisper, more grammatically correct or even more diplomatic. Machine-learning models need data sets on which to train and improve, and Grammarly employs textual data from its users in the training data it uses to improve its products and identify new ways people might use them.
“User inputs such as suggestion, accept or reject rates help us identify adjustments we may need to make in the product,” said Shane Collins, a Grammarly representative. “In some cases, we may store some text to help us improve the product, though we do not store all text transmitted through Grammarly.” Collins noted that the company removes user account data and personally identifiable information from text it uses to train its natural-language processing systems.
But Grammarly also sells an enterprise version of its writing tool that’s used by businesses and their employees inside text editors or even in apps such as Slack. And that data use by Grammarly’s business customers could create intellectual property concerns, said Jumi Kassim, a patent attorney at law firm Patterson Thuente, who also has a software engineering background.
“They are taking copies of anything you’re typing into their system,” Kassim said. For instance, when people use the Grammarly plugin while writing an email, “You’re giving them permission to take a copy of that,” she said.
After this story was published, Grammarly objected to that description, saying that it does not "store the entire content of documents in their original form."
“How much it is a big deal for companies depends on what data they are working with,” said Kassim. For instance, she said that while at a law firm like hers, software would have to go through a vetting process to ensure it does not compromise the security of confidential client information, “I’m sure there are lots of companies where they don’t care if you use Grammarly or not.”
Some tech companies say no
For some companies with employees writing about products or proprietary information, the data use is a red flag.
“We have had some pushback from clients,” said Elissa Ennis, head of Client Success at Enshored, which provides customer service, sales support and content-moderation services. The company offers its clients access to Grammarly to help them maintain brand consistency by setting up a customized Grammarly style guide.
Ennis said that customers such as software makers that are concerned about competition are especially leery of using Grammarly. “That’s only happened, really, I think two times where someone's been a pass on it. They did their due diligence on it. They talked to Grammarly, they talked to our IT about it,” Ennis told Protocol.
Software giant Microsoft reportedly said no to Grammarly too. GeekWire reported in 2019 that the company prohibits employees from using the writing tool. Microsoft declined to comment for this story.
It’s another example of shadow IT, when employees or teams purchase or use software for business purposes that companies should have processes in place to manage. “Any time your business is going to use a new piece of software, there should be someone vetting what software is being used and looking at the terms of service to see: Are they using my user data, and how?” said Kassim.
Intuit’s rigorous assessment
Intuit did just that before using Writer, a Grammarly rival, to help manage text for user interface elements like buttons and alerts content.
"Intuit, like I would imagine many Fortune 500 companies, is super-concerned about security,” said Tina O'Shea, its director of Content Design and Systems. “We had to go through a super-rigorous checking process … The security team had to vet it to make sure nothing was getting out.”
Writer calls itself a Grammarly alternative. On its website, it makes a point of highlighting the fact: “What you write will never make it into our machine learning models.”
Instead of employing user data to train its system, Writer’s CEO May Habib said the company uses a proprietary data set incorporating synthetic data, or information generated for the purpose of training its models.
Grammarly has run up against skepticism from critics who question the tradeoffs of handing over content in exchange for wordsmithing superiority, even when it comes to personal use. Amid the scrutiny, the company has altered its terms and conditions to limit how it employs user data such as text and documents.
Ultimately, said Kassim, businesses might want to consider paying for premium versions of tools that take less data.
“From a risk-management perspective, there are times when you can get away with things that have no cost to the business versus times it makes sense to pay money,” Kassim said. “There’s value in having a reduced risk.”
This story was updated to include additional comment from Grammarly.