Superhuman, the company that used to be Grammarly, just bought GPTZero, the startup that catches AI writing. The contradiction is the point. As the internet fills with machine-made text, proving something is human is becoming a product.
There is a neat irony at the centre of this deal. Superhuman’s biggest product helps people write with AI. It has just bought the startup best known for catching AI writing.
On Tuesday, Superhuman, the productivity firm formerly called Grammarly, agreed to acquire GPTZero. The companies did not disclose terms. PitchBook puts GPTZero’s value above $88m, Business Insider reported.
Founder Edward Tian says the company passed 19 million registered users and $30m in annual recurring revenue. It got there on total funding of just $13.5m, from a short backer list that includes Uncork Capital, Footwork, Jack Altman’s Alt Capital and Neo. The two cofounders and GPTZero’s 30 staff now join Superhuman to run a team focused on authenticity.
The pairing looks odd, and reporters said so fast. But the contradiction is roughly the plan. AI-written text now floods the web. Knowing what a human actually made is turning into a category of its own. Superhuman calls that category an “authenticity layer”, and it wants to own it.
A Princeton thesis that went viral
GPTZero is a tidy startup story too. Tian built the first version as a senior at Princeton in January 2023. The thesis project went viral almost overnight. He and cofounder Alex Cui, friends since high school, turned it into a profitable business on a shoestring. Tian, now 26, keeps the mission simple. “GPTZero started with the mission of preserving what’s human,” he said.
The product has grown well past an essay checker. It now spans hallucination detection, plagiarism checking and citation verification. A feature called AI Vision flags AI-generated content in social feeds in real time, across social media, email, publishing and reviews. Another, Replay, records how a document was written, keystroke by keystroke.
GPTZero says its detector holds a false-positive rate below 1% across 20 languages. Its hallucination checker has surfaced fabricated citations in academic and corporate work, including at a major AI research conference and at large consulting firms.
Why Superhuman wanted it
Superhuman already sells an AI detector inside Grammarly. One independent benchmark rates it top of its class. So why buy a rival? The public answer is glib: two detectors are better than one. The honest answer is reach.
The companies also argue the two halves fit. Detectors train on different datasets, so they flag different things. Pair GPTZero’s detection with Grammarly’s data on how 40 million people actually write, the pitch goes, and you get a fuller picture of what is human. “Together, we’re bringing the most trusted writing tool and the most trusted AI detector into one platform,” chief executive Shishir Mehrotra said.
Superhuman frames the goal as bigger than a yes-or-no verdict. Detection tells you whether AI wrote something. Tools like Grammarly’s Authorship and GPTZero’s Replay try to show how it was written, from first keystroke to final edit. That matters more as AI errors compound.
Hallucinated facts and invented citations seep into public documents, then train the next model. Catching them early, the pitch goes, keeps the rot out of the record.
Reach is the bigger prize. Superhuman claims 40 million daily users and a footprint across a million apps and websites. It will fold GPTZero into Superhuman Go, its AI assistant. Mehrotra likes to call a big acquirer a “trampoline” that helps a small company bounce higher. GPTZero is his fourth deal, after the workspace tool Coda, the Superhuman email app it took its name from, and the AI spreadsheet Rows.
The demand is real, and not only in classrooms. Education still drives about a third of Grammarly’s $700m-plus in annual revenue. Recruiters, consultants and journalists now want the same checks, Mehrotra said. The backdrop is stark. By one study Superhuman cites, AI now writes roughly half of all articles published online.
One tracker it points to predicts the figure reaches 100% within five years. Merriam-Webster just named “AI slop” its word of the year, and detection tools are spreading from schools to music-streaming platforms.
The catch, in two parts
Buy a detector and two hard truths remain. The first is that detection leaks. Researchers keep finding that popular detectors miss a large share of AI text, and a growing camp argues the tools were always going to lose the arms race. A detector that is right most of the time is still wrong often enough to ruin a student’s week, or a journalist’s.
The second is that Superhuman sits on both sides of the trade. Grammarly does not just flag AI text. It also generates it, then offers to “humanise” the result so it reads as though a person wrote it. The company recently drew complaints for AI feedback that imitated specific writers’ styles. Selling the poison and the antidote in one suite makes for a tidy business. It does not obviously make for a coherent mission.
None of that means the bet is wrong. The market may simply not care. Survey after survey shows that even as marketers pour AI into their work, audiences still want a human touch. If that holds, “made by a human” becomes a label worth paying for. Superhuman wants to be the one stamping it.
So the open question is whether authenticity is a durable product category or a passing anxiety. GPTZero bet its existence on the first answer. Superhuman has now made the same bet, at scale, while still selling the tools that make the bet necessary. Joining Superhuman, Tian said, lets GPTZero’s tools sit “at the exact moment someone needs them”, rather than as a separate step.
He framed the new goal as preserving critical thinking. Whether you can do that from inside the company helping to automate the writing is the thing worth watching.
Get the TNW newsletter
Get the most important tech news in your inbox each week.
