Inside GPTZero’s Office Hours with Founder Edward Tian
As the final event in GPTZero’s Teaching Responsibly with AI certificate program, the latest Office Hours session was led by GPTZero founder Edward Tian. We covered the motivations behind GPTZero, the links between media and AI literacy, as well as a couple of demos of GPTZero’s latest features.
Tian revealed the origins of GPTZero back to a moment in a classroom at Princeton, where he studied journalism and computer science. One of his formative experiences was taking a course with the legendary writer John McPhee.
“He's been teaching the same writing class for over 45 years,” Tian said. “He’s consistently loved going into the creative process of writing, the editing process improvement, selecting words as well. I was lucky to take this class.”
A year later, the two reunited to experiment with generative AI tools. Their question: could AI mimic McPhee’s style?
“He had a lot of questions of what AI could do and what AI couldn't do,” Tian recalled. “So you would ask him, AI, write in my style, and we would go into ChatGPT and be like, write in the style of John McPhee. And it could try… but it's not that good.”
The real turning point was when McPhee asked Tian whether any AI could recognize when a piece of writing was special and should be left untouched.
“He asked me, ‘Can AI tell you in terms of this essay and draft, that this is so good? Don't change anything about it.’” Tian said. “I was reflecting that there wasn't an AI tool that did that.” To which McPhee said: “Aha, that's a job of a great human editor.”
From that conversation came a lot of inspiration for Tian: “Part of what we're doing, and what hopefully we know we want our tools to do for students of the future, is not just to rewrite, generate, increase the quantity as well too in terms of words and texts on the internet, but also to preserve the really great quality, originality, and human moments of writing. And in doing so, highlight what's good, what's human… not just detect AI.”
Bridging understanding between educators and students
One of the key updates Tian shared was that, for the first time, GPTZero offered a free year of access to all teachers participating in the webinars. The only request was that they include GPTZero in their syllabi.
As Tian explained, many teachers start with a clear stance: they want to preserve space for human writing and avoid AI-generated work entirely. But as students start experimenting with new tools, the conversation shifts. Teachers are now asking: what’s the right balance? What does responsible AI use actually look like?
Tian shared the example of Eddie del Val in Oregon. In his first year using GPTZero, Eddie’s main focus was preventing AI use. But over time, his approach evolved. He began experimenting with new ways of integrating the tool, even adding it directly to the syllabus and encouraging every student to use it.
He introduced what he called the 80/20 rule: 80% human writing, 20% AI usage. For Tian, it was a refreshing move and a sign that teachers were becoming more comfortable setting their own boundaries, based on their classroom and their own personal knowledge of their students.
AI Literacy Is Media Literacy
Tian pointed out that the plagiarism conversation doesn’t just apply to student use of AI tools but also to the tools themselves. Many large language models are trained on scraped content from news articles, media sites, and journalistic work, often without attribution. These generative models, sometimes referred to as “stochastic parrots,” rely on massive datasets built from the internet, including the work of authors and journalists who are rarely credited.
He referenced the New York Times lawsuit against OpenAI, filed around eight months prior, as a sign of growing tension between major media outlets and AI developers. But beyond a few headlines, Tian noted that there’s still been no real resolution on attribution. For educators, this raises a media literacy issue that should shape how AI is introduced in classrooms.
Tian also spoke about hallucination, the tendency of AI to generate information that sounds deceptively confident and therefore somewhat credible, but isn’t actually accurate. He shared findings from an investigation GPTZero did with The New York Times, examining AI-generated books being sold on Amazon. What they uncovered was crazy, including guidebooks that seemed genuine but turned out to be entirely fabricated.
One such example was a travel guide to Uganda, which only revealed its made-up content after someone purchased it and realized the sites listed didn’t exist. Following the investigation, Amazon senior leadership reached out to GPTZero to discuss how to address the issue and prevent further reputational damage.
To help fight this broader problem, GPTZero has partnered with Wikipedia, giving editors access to the detection tool. As Tian explained, Wikipedia is a rare example of a human-led, community-driven platform yet it was also heavily scraped to train early AI models like ChatGPT.
Product Updates from GPTZero
Tian announced several recent updates. One of the most discussed is a new tag called Lightly Edited by AI, developed in direct response to teacher feedback. It recognizes the difference between fully AI-generated writing and student work that’s been passed through grammar helpers or lightly edited tools.
“This is based on a lot of feedback from teachers,” said Tian. “There’s a differentiation between AI-generated completely and written, but a lot go to grammar helper tools, whether it’s Grammarly or many others, for a light edit.”
Another update is a revamped writing feedback feature. “We built out a writing feedback feature and decided to categorize it into three steps. First, content - the overall, high-level flow and thought process. Two, clarity: how to make something clear in terms of how we're communicating, and only at the very end is grammar.”
This is due to all that Tian has heard from educators. “They used to care a lot about grammar edits and grammar corrections,” he said. “So if an essay had really poor grammar, they would penalize the students. Nowadays this teacher was sharing that they don't.” Instead they focus on the content and the thought process of the writing.
GPTZero has also launched a bibliography and sources checker, in response to concerns teachers raised last semester. “The pain point we've heard repeatedly from teachers... is that they're seeing bibliographies, not just the essays themselves, but bibliographies that are AI-generated, where when they click into the sources, links to articles that don't exist, or hallucinated, or referencing different parts of the text as well too,” Tian said. “Teachers don't have the time to click through every link during an exam season.”
The bibliography checker tool “goes and verifies the bibliography, searches on the internet what are hallucinated sources, what's credible, and sometimes what is tied directly to the content itself…because you might have links AI generated that are real, about, let's say, self-driving cars, but the article was about planes.”
Educators can test the tool directly on gptzero.me. “You’re welcome to try it out yourself,” Tian said. “Go to gptzero.me, on our ‘check bibliography’ tag as well too. You can try any of the example essays... you’ll see the quality, or put in one of your own essays.”
Another major feature, Authorship Replay, gives educators and students a clearer view of the writing process. “You can now see how a piece was written... being able to really jump through the different authors, because not only for the teachers, for individual students.”
To round off the session, Tian returned to the importance of process. “We’re not building an AI detector just for the sake of building an AI detector,” he said. “We have a research team that actively studies how this technology is evolving.”
He highlighted one of GPTZero’s research advisers who regularly works with their machine learning team and also leads AI model training for the LLaMA model at Meta. “It makes us proud that we have a lab committed to doing actual research in this space, and to exploring how it applies to academia. I think that’s a huge differentiator.”