A Princeton student created an app that outsmarts ChatGPT

Students looking to hand in AI-generated essays better think again.
Chris Young

Teachers have a valuable new tool against the relatively new problem of AI plagiarism.

Edward Tian, a 22-year-old senior at Princeton University, developed an app, called GPTZero, that detects whether a text was written by ChatGPT, a report by NPR reveals.

ChatGPT has gone viral in recent months partly due to concerns students could use it to generate essays in the blink of an eye, as well as a host of other fascinating and ethically concerning applications.

Now, teachers can use Tian's new app to fight back against the surge in AI-generated content.

GPTZero: A new tool designed to fight AI plagiarism

Tian, a computer science major at Princeton University, spent a large chunk of his winter break developing software he dubbed GPTZero. According to the student, the software is able to "quickly and efficiently" decipher whether an essay was written by a human or by ChatGPT.

Tian, who is also minoring in journalism, said he was motivated to create his bot to fight the increasing problem of AI plagiarism — whereby a human passes the work of an AI bot such as ChatGPT as their own.

In a tweet introducing ChatGPT, Tian wrote, "there's so much chatgpt hype going around. is this and that written by AI? We as humans deserve to know!"

Shortly after releasing GPTZero on January 2, the free beta service went offline, with Tian stating his free hosting service wasn't the influx of people who came to try out his software. Within a week of its launch, 30,000 people had already tried GPTZero. The student has also mentioned that many teachers have contacted him to tell him about their positive experiences with GPTZero.

How GPTZero catches out AI-generated essays

GPTZero works because it was trained on similar datasets as ChatGPT, meaning it can measure the comparative complexity — or "perplexity" reading — of a text. If a text is familiar to the GPTZero bot, it is more likely AI-generated.

Another reading GPTZero carries out is called "burstiness". This refers to the variation in sentence length and several other factors. While humans tend to write with variation, AI-generated texts are often more uniform.

Tian shared demonstration videos (above) in which he used GPTZero to analyze a story published in The New Yorker alongside an article that ChatGPT knowingly wrote. GPTZero successfully distinguished the AI-written text from the New Yorker piece.

Though Tian's work has great potential for helping teachers catch out cheating students, it isn't 100 percent foolproof. It's worth noting that an artist was recently unfairly banned from a subreddit after being accused of using AI.

Still, Tian sees his work as the beginning of a push for increased AI transparency. On Twitter, he wrote that "for so long, AI has been a black box where we really don't know what's going on inside," he said. "And with GPTZero, I wanted to start pushing back and fighting against that."