Tech companies are drowning in words. Think about all the code, website content, social media copy and internal communications piling up inside of them. Who monitors these words to ensure they fit the company’s style and voice?
For a growing number of companies, the answer is AI. Intuit, the parent company of TurboTax, Credit Karma and Mailchimp among others, uses an AI assistant called Writer. Outsourcing company Enshored uses Grammarly Business, the enterprise version of the popular writing helper. The tools work anywhere there’s a text editor, from higher-stakes blog posts to informal Slack messages.
The idea of using AI to help your writing isn’t new – Grammarly has been around since 2009. But remote work has led to the growth of asynchronous, often written communication. Tech workplaces are thinking more about the best ways to communicate, and writing quality is an important consideration. When we’re not in-person, we often rely on written words to best convey our tone and meaning.
“That message is carrying a lot more on its shoulders than it did before,” said Rahul Roy-Chowdhury, Grammarly’s head of Product.
How do AI writing tools work?
Tools like Grammarly and Writer use machine learning and natural-language processing technologies to improve writing. While both take on more nuanced tasks such as helping ensure language is inclusive, at a very basic level, the systems aim to improve writing by correcting things like spelling, punctuation and grammatical errors. To do that, they attempt to translate examples of incorrect language into correct versions.
Timo Mertens, head of Machine Learning and NLP Products at Grammarly, said the company starts off its process with human linguists who define a problem – say incorrect use of a preposition in a sentence. Those linguists gather data including anonymized information gleaned when Grammarly users write using the tool. The linguists use that data to train rules-based models for determining ways the system should translate incorrect language into correct language.
To augment and refine that human-centric process, the company combines an automated machine-learning process with what its linguists built, Mertens said.
Writer also works by translating something that’s incorrectly written into a corrected version. However, rather than using real-life user information, the company trains its machine-learning models using a giant set of synthetic data, or examples of writing created specifically to train algorithmic models.
“Whereas there is lots of data online that's the same document in English and French, for example, which can be used for [language translation] training data, there’s nothing like that for writing,” said Writer’s CEO May Habib.
How tech companies use AI tools
Intuit first started using Writer to help manage the company’s sprawling UI text strings: the words on a website’s buttons, labels and headings, for example. Once developers have finished a product, it’s extremely hard for non-developers to change UI strings on their own — even if it’s as simple as deleting a space or period. Writer, however, is able to easily go into these strings and edit them. Twitter uses Writer for this purpose as well.
The other main use case is improving writing quality. This means abiding by the user company’s up-to-date style guide, but also means using more inclusive language. As Intuit’s director of Content Design and Systems, Tina O’Shea has internalized the company’s style rules. But that’s not the case for everyone. “There are many people outside my writing team who are writing content, sometimes customer-facing content, but everybody's writing a gazillion internal communications,” O’Shea said.
Photo: Grammarly
Tools like Writer and Grammarly bring style rules directly to users. Elissa Ennis, head of Client Success at Enshored, said Grammarly helps maintain brand consistency within Enshored, but also among its various partners. Enshored works with over 50 companies, and offers each of them their own Grammarly style guide. Not all of them say yes (those with proprietary software are concerned about Grammarly’s data-training systems). But when they do, Enshored’s outsourced workers can easily adopt the client company’s tone.
“We work with a comic-book-style business that wants everyone to be called dude to maintain that casual and edgy atmosphere,” Ennis said.
The AI can read into your writing tone as well. The Grammarly assistant will pop up while you write an email, informing you that the message sounds “curious” or “informal.” Ennis said this feature has helped her nail her communication approach. You might want to sound friendly and approachable when introducing yourself to a new hire. But in an all-staff email, you might prefer a formal and authoritative tone.
“Those subtle nuances and language help get your point across and help maintain good communication and relationships with your co-workers,” Ennis said.
What does this mean for communication in the workplace?
Features like Grammarly’s “tone detector” or Writer’s “healthy communication” ideally lead to more thoughtful interactions. Like Roy-Chowdhury said, messages carry more weight when you barely ever see your co-workers in person. Habib agrees. “When it is 2D, plain-text only, you can basically assume that the recipient is going to interpret what you said in the worst possible way,” Habib said.
An easy way to be more thoughtful is to eliminate outdated or exclusionary language. Writer made sure to bake inclusive terms into the product from the beginning, Habib said. For example, avoiding pronouns in job descriptions. Customers can then customize on top of that. Intuit took advantage of this, inputting antiracist and readability style guidelines into Writer’s API. A developer at Intuit then used the API to create an inclusivity Slack bot, which offers private suggestions on improving employees’ Slack messages.
O’Shea said Intuit employees had a range of opinions on the inclusivity bot. Some appreciated the opportunity to stop and think about their words. One response: “Organic teachable moments. So powerful!” Another person said they used the word “insane,” and was prompted to think about mental health and ableism. “I honestly felt a bit called out, but took a step back, and now I’m thankful that I have a safe sandbox to learn!” they said.
Others found the bot more frustrating than helpful. They found it excessively pushed back on words. At least one person was concerned their manager might use the bot to analyze performance management. To clarify this was not the case, the bot now introduces itself with the phrase “no one can read this except for you.”
Providing unhelpful or off-putting suggestions is a very real concern for Grammarly and Writer. Users are easily spooked by suggestions that come off as patronizing or condescending. In addition to instructing users on their tone, AI tools need to consider their own tone as well. No one wants to wade through a million somewhat annoying AI suggestions while writing off a quick Slack message. Balance is key.
At their best, AI writing tools make communication easier across all demographics: country, age, culture, job title. They can help eliminate corporate jargon, or adjust to language dialects. The workforce is more digital and dispersed than ever. We’re still exploring the best ways to talk to each other. Writer and Grammarly think they’re best positioned to help us tackle the sea of words.