In June, Superhuman earned GPZeroa New York-based startup best known for building one of the most popular tools for spotting AI in student writing. Co-founders and her new team have since joined Superhuman’s newly formed “authenticity” group, where they’re working to integrate detection capabilities across the company’s products, most notably Superhuman Go, an AI assistant designed to operate on everything a user does online.
The deal marks the latest step in a rapid transformation led by the CEO Shishir Mehrotrawho took the helm in early 2025 after the company, then called Grammaticallywon his productivity startup, Code. That summer, Grammarly acquired Superhuman Mail, an email tool for power users built around the idea of ”inbox zero”. In October 2025, Mehrotra renamed the combined company Superhuman, signaling an ambition to build a suite of AI tools beyond the typing assistant.
Before founding Coda in 2014, Mehrotra built and launched products in to YouTube AND Microsoft. He is repeating a popular Big Tech playbook—Google it was done Alphabet in 2015; Facebook it was done Meta in 2021-where a popular product evolves into something broader.
“There have been plenty of cases where this has been done in a way that preserves the core brand,,” Mehrotra told the Observer at the time.
Grammarly, still Superhuman’s most popular product, is used by 40 million people every day. Students rely on it to refine essays; editors use it to catch typos and enforce consistency of style—sometimes across sprawling internal regulations that can run hundreds of pages.
The use of AI in writing—perhaps one of the most humane forms of work—has been controversial from the beginning: How much use of AI is too much? Is AI best used to generate a first draft or do the final polish? And increasingly, can anyone reliably tell the difference?
Mehrotra has had to navigate those questions publicly. Earlier this year, the company faced backlash to its “Expert Review” feature.which generated suggestions in the style of popular writers and led to a class-action lawsuit over consent.
“There’s a clear line, depending on the task, between things you should use AI for and things you shouldn’t,” he told the Observer in an interview earlier this month. “In the case of a writer, it can be a great tool for doing research or understanding a topic. But if it sneaks into the actual writing and takes away from the person doing the reporting, it reflects badly on both the writer and the media.”
Teachers and editors have long complained that AI detection tools are unreliable, while many AI researchers argue the problem is fundamentally intractable. In any case, Mehrotra is already looking at a bigger picture. “Think of Grammarly as a great English teacher sitting on your shoulder all the time, helping you wherever you work. But it’s actually a massive underuse of the infrastructure we’ve built,” he said. “Cessical infrastructure is the ability to bring pervasive AI to where you work.”
This idea underpins the company’s rebranding and, not long after, its launch of Superhuman Go, which extends Grammarly’s real-time help to a more general-purpose agent system.
“We’re enabling anyone to build an agent that works just like Grammarly, but does things far beyond grammar,” the CEO explained. “For a salesperson, that might mean having a sales coach by your side, warning you that you’re about to recommend the wrong product while emailing a customer. For a support agent, it might mean a support coach reminding you that a customer had a big stretch yesterday and suggesting you mention it.”
Looking for human qualities in a world saturated with AI
Superhuman’s multiple departments — from engineering to sales to customer support — use nearly every AI tool on the market, Mehrotra said. But when the information reaches the level of the CEO, he faces his own test in drawing the line in the use of AI.
“I often tell people: use AI to help you form your opinions, but not to write your briefings and reports,” he said. “When people send me something that’s clearly written by an AI, I’ll often tell them, ‘I’d rather just have the prompt.’ I’d rather see your unique, raw insights and start from there, rather than read a polished AI output.”
The prevalence of AI has also changed the hiring process. Technology leaders like Bill Gates predicted early on in the generative AI boom that incorporating AI into work would be like having a personal digital assistant around the clock. And that has changed what managers look for in job candidates, even for entry-level positions.
“AI has added another bar to almost every role; we’re turning everyone into a manager very early in their experience. You’ll have a team working under you from day one, which is completely different to how it used to work in the pre-AI world,” Mehrotra said. “Suddenly, we have to look for very similar skills (as managers) even for entry-level candidates. Your ability to work with AI to produce the right result is incredibly similar to your ability to work with a team of human employees.”
What remains is the search for uniquely human qualities. Mehrotra and his co-founder Coda, Matt Hudsonhave evangelized the idea of applying the “eigenquestion” discipline to decision-making – a borrowing from mathematics that reformulates prioritization.
“We often find ourselves in situations where we have ten questions to answer and we have to prioritize them,” he explained. “People usually rank them by severity or urgency. But if you use the self-questioning test, you look at that list and ask, ‘If I answer this one question, how many other questions on the list does it answer?’ It could be question number six on a list of ten.”
This way of thinking is especially important in making strategic decisions. As AI becomes highly capable of generating reliable answers to everything, the scarce resource becomes determining which problems are worth solving and which results to trust.
“When I think about the misuses of AI, I think that people often try to replace human knowledge with AI. When they do that, they end up, at best, embarrassing themselves, and at worst, actually producing the wrong results,” Mehrotra said. “One of the things that is uniquely human is judging with very little information. After all, knowing what problem to solve also seems, at least for now, uniquely human.”
