How the 23-year-old founder of GPTZero turned his student anti-cheating tool into a leading startup that detects AI-generated text online

• Edward Tian first built GPTZero as his senior thesis project at Princeton during his winter break.
• Since then, he's turned it into a full-fledged startup that's raised $3.5 million in VC funding.
• He's expanded the company's mission from anti-cheating detection to verifying "humanity" online.

Last year, Edward Tian was looking for an exciting new project to cap off his senior thesis in machine learning at Princeton University.

The then 22-year old computer science major wanted to create a tool that combined his interest in AI and his minor in journalism.

Over the course of the holiday break last year, Tian built what would become the main tool for his startup, GPTZero. The technology Tian created uses machine learning to scan a block of text and identify whether or not the text was written by a human or with an AI tool, like OpenAI's ChatGPT or Google's Bard.

He launched it online in January, and it instantly became a smash hit, with 30,000 users signing up in just the first week, he told NPR at the time.

Educators, in particular, found the tool incredibly useful to detect against student cheating on essays and other writing assignments, since other plagiarism-detection scanners like the website TurnItIn.com were not yet capable of catching AI-generated text. They became some of the earliest adopters of GPTZero, Tian explained.

"You have to imagine, this is before TurnItIn even knew what AI detection was, and OpenAI wasn't considering this at all either," Tian said.

Tian had been influenced by some of the work he had done the year before at the BBC. He had taken a gap year to work at the British broadcaster to research solutions for stopping disinformation. "The target audience for the tool was myself, actually," he said, "because we were looking a lot into disinformation on the internet, and this tool would have been really ideal."

As customer and media interest grew in GPTZero, Tian realized he had built something that could be a full-fledged startup. Even though he had a journalism job lined up after college, the buzz around GPTZero was too great. He decided to team up with his friend Alex Cui, who was working on getting his PhD in computer science at the University of Toronto, to focus on GPTZero full-time.

Cui dropped out of his program and became GPTZero's CTO. This spring, the two raised a $3.5 million seed round led by Uncork Capital and Neo. Lattice CEO and founder Jack Altman's firm, Altman Capital; the CEO of Stability AI, Emad Mostaque; Reuters CEO Tom Glozer; and former CEO of the New York Times, Mark Thompson also participated in the round.

Tian said he sealed the deal on his fundraise over the course of his senior spring break.

In the seven short months since its launch, Tian and Cui have a team of 12 employees around the globe — many of whom are contract developers in Uganda who Tian knows personally, and a machine learning engineer based in Kenya. And GPTZero was recently recommended by Greylock's Asheem Chandna for Insider's list of the most promising startups of 2023.

GPTZero detects whether or not a chunk of text is AI-generated by looking for two main qualities: "burstiness," which is the variability of word choice and sentence structure in a given block of text, and "perplexity," which is how "perplexing" a graf of text is to the AI model that's scanning it. It's not perfect by any means, and can get tripped up on non-native English speakers' writing, since oftentimes the syntax can be more repetitive.

And Tian is clear he doesn't want the tool to become the only tool teachers use when determining if a student cheated on an assignment. While more AI detectors have popped up since the launch of GPTZero, they've been riddled with errors. TurnItIn's own tool was criticized for its high number of "false positives" in detecting AI generated text in assignments, and OpenAI's own tool had so many inaccuracies it has since been shut down,

Other AI detection tools have gotten better, however. AI juggernaut Hugging Face has started offering its own AI detector tool, and others like Originality.ai have popped up that cater to specific industries, such as content marketing.

Tian and the GPTZero team acknowledge that errors can happen with GPTZero as well, but that focusing on "human" detection can be one way around it. The startup recently launched its newest tool "Origin," a plugin that can be integrated into Microsoft Word and track a student's writing process to verify that it was really produced by that person.

"Our model has been able to improve because we've been building on top of new research on how to improve AI detection models, when everybody else has been copying and using the current AI models," Tian said.

Going forward, Tian and Cui have expanded the scope of GPTZero beyond plagiarism detection to be a broader AI-detection utility. Currently, GPTZero is free to use for up to 5,000 characters of text, and pricing for enterprises ranges from $9.99 a month to $19,99 a month, with an additional $14.99 a month fee for developer API access.

"The idea here was to always build out the best technology," Tian said, "and then give it to as many industries that are being affected by AI as possible, including education, finance, journalism, cybersecurity and everything beyond."