Alex Cui (left), CTO and head of R&D at GPTZero, dropped out of his PhD to join CEO Edward Tian (right) to build AI detection software GPTZero and bring transparency to machine learning.
Edward Tian
CEO Edward Tian and his team of 10 machine learning engineers want to “preserve the human stuff” by being able to identify what machines generate.
In December 2022, Edward Tian prototyped GPTZero, a ChatGPT detection tool, from his Princeton University dorm room. As the world became more and more obsessed with his ChatGPT, Tian was starting to worry that his AI-generated content disguised as human-written articles and essays would soon become harder to find. The 22-year-old spent his winter break building his GPTZero to ease this. Then he used spring break to secure seed funding for it.
“It’s like opening Pandora’s box,” Tian said of ChatGPT in an interview in January. Forbes. “It’s suddenly thrust into the world and has a lot of potential for exploitation.”
GPTZero, which has amassed 1.2 million users in the past five months, raised a $3.5 million seed round on Monday from early-stage ventures Uncork Capital and Neo, Altman Capital and Stability AI CEO Emad Mostaque. announced. Reuters CEO Tom Glocer and former New York Times CEO Mark Thompson also participated in the round.
“I think it was important for people to bet on independent third parties to build protections for their technology,” Tian said last week, adding that the proliferation of large-scale language models is now “polluting the internet.” He added that there were growing concerns.
GPTZero scans text using its own large-scale language model, an “ensemble” of other open-source AI models, trained on both human- and AI-generated text, such as news articles and questions and answers. will be The tool computes and predicts the probability of words in AI-generated sentences by learning from existing generative AI models. GPTZero also analyzes written patterns using syntax and sentence length to identify machine-generated text. Tian admits there is always room for improvement, but it does a good job. Because the chatbot’s output depends on the data and prompts it is given, the tool may flag human content as produced by her AI and vice versa. As such, GPTZero uses both machine learning and human work to recreate content from various AI programs and evolve to more accurately detect AI-generated text.
GPTZero can detect text from various AI models, including Google’s LaMDa (also known as Bard), Facebook’s LLaMa, and OpenAI’s GPT-3 and GPT-4. After launching, GPTZero partnered with educational institutions such as Canvas and K16 Solutions to collect training data for their AI detection tools. An educator found it helpful for students to detect essays that used her ChatGPT.
But there are many other applications as well. The startup plans to market its services to corporate customers in publishing, social media, trust and security. Tian said governments are also potential customers and believe AI detection tools can be used to create policies. The company’s $9.99 monthly plan already has 3,000 paying subscribers.
The startup will soon roll out a browser plugin called Origin to assess published text for accuracy and factual provenance. With Origin, the startup hopes to support media literacy in the age of artificial intelligence by fact-checking AI-generated content and verifying citations of AI-generated content. “It’s more than finding students abusing tools, it’s not even detecting AI,” Tian explains. “It’s about preserving what is human.”
