Preserving What’s Human: Our Journey With GPTZero James Kim

When ChatGPT launched in November 2022, a prevailing sentiment in education was alarm and despair. Students would cheat. Essays are over. What's the point of homework, or tests?

We focused on two other ideas.

First and foremost, we believed that prohibition was never going to be the right posture toward AI in the classroom. Ben Thompson argued this early and well: the technology that looked like the end of homework might be the beginning of a better way to teach it. Used deliberately, with proper guidance from a teacher, AI could be complementary to learning. Clearly the genie was out of the bottle. But could it be channeled to help learners understand how to wield a new, powerful tool for a new world?

The second: detecting AI-generated text is a much harder technical problem than it looks. The same prompt yields different output every time, and each new model writes a little more like a person. Detection is never a fixed target; it has to be re-solved with every release.

These were overlapping ideas but they rarely converged in the same team. Builders technical enough to chase the detection problem generally have no patience for pedagogy, and the people steeped in pedagogy usually aren't technical enough to train detection models.

Edward Tian and Alex Cui, co-founders of GPTZero, were the exception. Alex brought the machine learning depth from his time as an undergraduate at Caltech and a Master's student at the University of Toronto. Edward also had technical chops, as a computer science major at Princeton with a journalism minor, and a stint at BBC working on an investigation team tracking disinformation spread.

More importantly, they had the foresight and vision to articulate why this work mattered. Their framing (which I've never improved on and still use) was that recognizing AI is not the same as catching plagiarism. Detection should be the common ground that lets teachers and students navigate AI together, not a tripwire.

We got to know Edward and Alex in the Reach Way - working closely alongside them even before funding. With our deep experience in edtech, we shared how the company could think about its approach to K-12 and higher ed go-to-market. Today, GPTZero is used by more than 19 million users, including students and educators at schools and universities across the world.

As a board observer, I was treated to a front-row seat to the company's velocity. A tight-knit hacker culture (which Edward calls "Group Project Energy") is infused throughout the company. GPTZero ships in tight pods of engineering, design, and ML, faster than the frontier moves: detection coverage for each major new LLM usually lands within about two weeks of its release.

The product compounded accordingly: an editor, writing-replay and origin reports, LMS integrations, an AI grader, and AI Vision (which flags AI-generated text as you scroll your social media feed). There's a live tracker of how much of the internet is machine-written. They've also built the first detector I'm aware of that could catch fabricated, AI-generated citations, something that has been making research instructors miserable. (They've caught some high profile misuses.) They built all of this while staying capital-efficient, which in this vintage of company is its own kind of statement.

My favorite evidence that the thesis played out isn't in any metric or stat, though. It's a college writing professor who started out using GPTZero to catch students using AI, and a year later rewrote his syllabus to require them to use it - as a way to find the line where the work was still authentically theirs. That's the whole idea, realized in one classroom: not surveillance, but calibration.

We are incredibly proud that GPTZero will be joining Superhuman, the company formerly known as Grammarly. Edward describes the move as a "trampoline" - a larger company whose role is to help a smaller one jump higher rather than absorb it. The mission doesn't change: preserve what's human, now that half of all articles being published on the internet are written by AI. It just gets a far bigger surface to run on.

Backing Edward, Alex, and this team has been one of the real privileges of my time at Reach. I can't wait to see how high the trampoline takes them.

P.S. The GPTZero plugin estimates there is a 78% probability that this was human-written and polished with AI, which sounds about right.