Amongst all of the young AI startups being ruthlessly pursued by VCs as of late, GPTZero has already grown into profitability in its first yr and a half of life, generating thousands and thousands in revenue. Founded by 24-year-old Edward Tian and 26-year-old Alex Cui, who’ve been friends since highschool, GPTZero offers a detection tool that helps discover whether a chunk of content was AI generated.
The founders have chosen to take a $10 million “preemptive” Series A led by Footwork co-founder Nikhil Basu Trivedi, the team has exclusively told TechCrunch. (“Preemptive” is VC-speak for when an investor nabs a deal before the founders were trying to boost.)
This is kind of the coup for Basu Trivedi. GPTZero has been watched by top VC firms practically since Tian launched an initial version as an online app in December 2022, and 30,000 people immediately swarmed it, crashing its Streamlit-hosted website. (Adrien Treuille, Streamlit’s co-founder, who sold to Snowflake for $800 million, later became an angel investor, Tian says.) The corporate formally launched in January 2023.
Throughout 2024 thus far, as its customer base has grown, the young founders were fielding 4 to 5 calls from VCs per week, they said.
GPTZero grew 500% in ARR within the last six months, the founders told TechCrunch, adding that its user base has grown from 1 million to 4 million within the last 12 months. This makes it one among the fast-growing consumer apps of the yr, by some measures.
The corporate has been profitable for the last several months, they said, adding that they’ve more cash within the bank than the whole raised within the lifetime of the corporate. To place a number to that: greater than $13 million between its $3.5 million seed and the brand new $10 million.
And the expansion continues. Users and revenue have “greater than doubled, perhaps even tripled, since January,” Basu Trivedi said. While they didn’t comment on valuation, based on a typical 20% Series A round, the deal has valued the corporate somewhere across the $50 million mark pre-money. Other investors within the round include education-focused (and women-led) Reach Capital; Jack Altman’s Alt Capital; Uncork Capital (Jeff Clavier’s fund); and Neo (Ali Partovi’s fund).
How the VC won the deal
Basu Trivedi, a Princeton alumni, won the lead on this deal by playing the long game. He met Tian in 2022, before GPTZero craziness, during an annual event where a small group of Princeton students visit Silicon Valley corporations. Basu Trivedi all the time takes the group on a hike of the Stanford Dish.
Tian developed GPTZero while he was studying computer science, natural language processing and journalism on the Ivy League school. During internships for the BBC, and later at The Latest York Times, he wrote code that helped journalists discover AI-generated content.
After the wild response his initial web app got, Tian reached out to his buddy, Cui, for help. Cui has a master’s in machine learning from the University of Toronto and dropped out of his doctorate program to turn out to be a co-founder.
The 2 rewrote the app into its current stand-alone platform and raised the $3.5 million in seed after reaching about 1.5 million users in its first five months. This got here mostly from angel investors like Tom Glocer, former CEO of Reuters; Russ Salakhutdinov, Carnegie Mellon University professor and ex-director of AI research at Apple (after he sold his startup, Perceptual Machines, to Apple in 2016); and Mark Thompson, CNN’s CEO and former Latest York Times CEO.
Basu Trivedi saw how GPTZero was gaining press and impressive angels — and heard the rumblings about it among the many VC scuttlebutt. As a seed investor who backed corporations like Canva, ClassDojo and Frame.io, he knew a hot company when he saw one.
He texted Tian in January 2023 to examine in. He wooed the founders together with his network and product know-how from his fast-growth corporations like Canva, and with the background of his fund’s co-founder, Mike Smith, former COO of Stitch Fix and Walmart.
Investors with each product and operations experience were what the 2 20-something founders were “craving, especially as Alex and I are learning tips on how to construct a giant company,” Tian said.
To prove the purpose, shortly after they closed the round, Footwork organized a networking event with AI leaders, including Jack Altman (brother to OpenAI’s Sam Altman, who joined the Seed A round), who was also a university classmate of Basu Trivedi, and Nvidia founder CEO Jensen Huang.
“A giant data advantage“
GPTZero is way from the one company working to discover AI-generated content. Others include AI Writing Check, Copyleaks, GPT Radar, CatchGPT and Originality.ai.
But many within the AI-detection industry have abysmal accuracy, researchers find. A lot in order that OpenAI, which was pressured by AI-industry paranoia into launching its own AI detector at first of 2023, shut the tool down about seven months later in July, after it was widely criticized for a way poorly it worked.
Interestingly, when TechCrunch’s Kyle Wiggers did his own experiment with these tools, all of them flunked except GPTZero.
Naturally, GPTZero has its own benchmarks, particularly through a partnership with Penn State researchers, that help it make its case that its tech works well, despite the industry’s general fame.
Cui says GPTZero is more accurate since it has access to more data and has built its own LLM models using probably the most advanced open source tools, which it won’t disclose.
“Now we have a giant data advantage. Now we have thousands and thousands of examples of text that’s human versus AI,” Cui said. “We’ve also combined this with a number of the best-in-class models and deep learning. We’re actually using language models to detect language models.”
While the startup could also be best known for helping teachers detect AI-generated student work (in October, GPTZero landed an agreement with the American Federation of Teachers), its customer base has expanded. It now includes government procurement agencies, grant-writing organizations, hiring managers, and — especially interesting — AI training data labelers.
It seems, using AI-generated data for AI training “causes model collapse,” Tian says, because teaching a model using fabricated examples isn’t the most effective approach to get it to operate in the actual world.
Naturally, the young founders have a more grandiose long-term vision. They need to create a brand new, independent layer of the web that performs accountability, ensuring that human and AI content is correctly attributed.
To that end, the team is currently working on AI hallucination detection. Hallucinations, where the AI presents AI-generated fiction as if it were fact, are the bane of the GenAI industry. The corporate’s first step toward addressing it is a newly available free AI text copyright check for LLM training datasets. This can help them generate the training data for broader hallucination detection.
“We’re just attempting to avoid a world where the whole web is AI-generated content,” Tian said. “A web where everybody uses AI doesn’t preserve the chance for people to proceed contributing creative and original content.”
