College student made app that exposes AI-written essays

ChatGPT’s artificial intelligence generated dialogue has gotten pretty sophisticated — to the point where it can write convincing sounding essays. So Edward Tian, a computer science student at Princeton, built an app called GPTZero that can “quickly and efficiently” label whether an essay was written by a person or ChatGPT.

In a series of recent tweets, Tian provided examples of GPTZero in progress; the app determined John McPhee’s New Yorker essay “Frame of Reference” to be written by a person, and a LinkedIn post to be created by a bot. On Twitter, he said he created the app over the holidays, and was motivated by the increasing possibility of AI plagiarism.

On Jan. 3, Tian tweeted that GPTZero wasn’t working, likely due to a larger than anticipated amount of web traffic. In a Substack newsletter Tian published today, he said that more than 10,000 people had tested out the publicly available version of GPTZero on Steamlit. (At time of writing, both gptzero.me and the Streamlit version are showing errors, likely due to volume of traffic.) In the newsletter, Tian said he updated the GPTZero model to “significantly reduce the rate of false positives and improve output results.”

GPTZero uses “perplexity” and “burstiness” to determine whether a passage was written by a bot. Perplexity is how random the text is in a sentence, and whether the way a sentence is constructed is unusual or surprising to the app. Burstiness compares these sentences to one another, determining their samey-ness. Human writing has more burstiness — which is to say, we tend to write with more sentence variation.

Concerns about plagiarism have abounded since OpenAI launched ChatGPT on Nov. 30, 2022. More than a million people used it within five days post launch. The AI-powered software can write basic essays and emulate the style of established writers. You can direct ChatGPT to copy Shakespeare’s voice, for example, or write in the style of a New Yorker essayist. There are snags in execution, but results are recognizably in the right style. It’s not hard to get the AI to write a high school English-style essay, and to find the result pretty indistinguishable from an assignment written by a student. That said, there are still limitations to what it can do. It’s easily baffled by riddles, and sometimes just makes up facts. StackOverflow also banned any ChatGPT-generated coding feedback, thanks to the frequency of errors.

In December, OpenAI said it would “watermark” ChatGPT output, in order to combat plagiarism.

In his newsletter, Tian said he’s working on more updates to GPTZero, including “improving the model capabilities, and scaling the app out fully.”