Last winter, the first version of GPTZero, software built by Edward Tian ’23, went viral. GPTZero, created by Tian out of his senior thesis work, is designed to detect text written by artificial intelligence. Since then, large language models like ChatGPT have progressed.
Tian’s product has developed also. After GPTZero was released, Tian founded a start-up of the same name. Devoting his time to the start-up full-time after graduating, GPTZero now has twelve staff, according to its website, and recently took on two interns from Princeton. A preprint of a study on AI detectors showed that, by some measures, GPTZero was the best-performing software out of several competitors at detecting AI text.
Yet the study also backed up concerns that AI detectors can in general be fooled and are likely to have false positives, which can lead to undeserved disciplinary action for students. To address false positives, GPTZero is investing in a feature which students would use to track their writing and prove AI was not used.
According to Tian, GPTZero constantly undergoes updates in service of the dual goal of creating software that can reliably discern instances of AI-abetted plagiarism and provide authentication to certify the originality of human text.
The original version of GPTZero measured perplexity — the randomness of the word choice and construction of a given sentence — and burstiness — a comparison of perplexity across sentences.
Two weeks later, on Jan. 15, Tian’s group released an updated version of GPTZero called GPTZeroX, with new features such as highlights for particular phrases and sentences that exhibit a high probability of having been produced by AI.
In February, just over a month after Tian tweeted out the beta version, the group launched GPTZero’s API, with the central mission of making the software accessible. According to the most recent Substack post by the group on June 22, several universities have harnessed the abilities of GPTZero within academia: “Researchers, ranging from Stanford, Harvard, Berkeley, etc. have also integrated and tested our APIs, often validating GPTZero as one of the most robust and accurate AI detection models.”
The Stanford study that noted GPTZero as one of seven leading AI detectors also warned that AI plagiarism detection software is liable to convey biases against ESL (English as a Second Language) writers.
To combat concerns that software may falsely accuse students of using AI, Tian and the team at GPTZero have focused on helping students prove that their work was not AI generated.
Tian’s group has released GPTZero’s v2 API, which incorporates a number of new features. The software’s human authentication feature, called “Origin,” was specifically designed with students in mind. Users can harness this newest functionality of GPTZero to document their entire writing process, thereby providing a “human-print” proof in video format demonstrating that AI was not employed to produce written text.
“What we’re doing now is AI detection, but we’re also doing a verification tool where you can write whatever and then on Google Docs you can get a writing report of the actual writing process. And the idea here is that the detection is like a scan where the algorithm predicts whether this [the written prompt] is AI or not,” Tian explained in an interview with the ‘Prince.’
In an era where educators have expressed growing concerns about AI facilitating plagiarism, Tian hopes this new feature will empower students and teachers alike to embrace AI technology.
Among GPTZero’s proponents is Tripp Jones, a general partner at Uncork Capital, a venture capital firm. In a blogpost to Medium, Jones hailed GPTZero as a revolutionary new tool to counteract the dissemination of fabricated information bound to be unleashed by generative AI technology.
“This dual-functionality could be a game-changer in the fight against misinformation,” Jones wrote.
GPTZero’s v2 API launch notes that GPTZero undergoes constant updates, with the most recent Substack post stating that the software “evolves rapidly, and is updated nearly every week!”
The software capitalizes on deep learning algorithms to optimize its plagiarism detection accuracy while simultaneously interweaving elements such as an internet text search functionality that browses troves of web data to authenticate text originality. In addition, the software makes use of “GPTZero Shield,” a novel feature that buffers the software against tools that may seek to “exploit AI detectors.”
The post emphasizes that the model’s reliance on deep learning has contributed in no small part to GPTZero’s accuracy, with training from “massive text corpuses from the web, education datasets from our partners and also our own synthetic AI datasets generated from a range of language models.”
“Gigantic datasets actually analyze the language in text and it gets better as we have a larger amount of language data. And it gets better as we finetune a model for specific types of writing, whether it’s student writing or other types of writing,” Tian said.
AI detection software has been criticized for the false identification of text. For instance, TurnItIn, an AI writing-detection tool, recently revealed a proliferation of false positive results in its evaluation of written content, indicating that the system inaccurately flagged human-produced texts as AI-generated at a far higher rate than the company originally disclosed.
Regarding GPTZero’s propensity for creating false positives or negatives, Tian noted that his company’s software has accounted for these potential flaws through the deployment of appropriate software guardrails. As a result, GPTZero boasts a far higher detection accuracy rate relative to comparable models.
“What we’ve been really unique in doing compared to a lot of other folks is being very conservative on the false positive side,” said Tian. “How we measure accuracy for our model was to tune things to be 99 percent accurate on a testing dataset for detecting if AI is AI, in terms of minimizing both false positives and false negatives.”
Tian also said that GPTZero is retaining its Princeton ties by recruiting University students as summer interns.
One intern, Maggie Wang ’26 wrote in a statement to the ‘Prince’ that she appreciates that “GPTZero is diligently working on adaptive solutions that acknowledge the transformative influence of AI on education, while upholding the innate value of the human voice, the authenticity of unique perspectives, and each individual's personality.”
“The team’s unwavering enthusiasm and ability to fearlessly confront [difficult and complicated] matters head-on inspire me and make me proud as a new member of the team,” Wang wrote.
Another intern, Jin Schofield ’26 echoed these sentiments, noting that she was excited by the prospect of GPTZero transforming into a potent educational tool.
“As GPTZero evolves, I am really excited to see how we can help improve transparency between students and teachers regarding AI use in the classroom,” Schofield wrote in an email to the ‘Prince.’
In the future, Tian said that he intends for GPTZero to become the blueprint for using AI to enhance the writing process in a world rapidly being shaped by technological advancements.
“I think the ultimate goal here is to build not just the detection platform for the Internet but the writing platform of the future in a world where people are writing with some level of AI involvement or not,” said Tian.
Tian served as a senior News writer for the ‘Prince’ during his time as an undergraduate student.
Amy Ciceu is a senior News writer for the ‘Prince’.
Please direct any corrections requests to corrections[at]dailyprincetonian.com.