Edward Tian Creates Generative AI Detector GPTZero:

Greg Rosalsky and Emma Peaslee of NPR tell us that “This 22-year-old is trying to save us from ChatGPT before it changes writing forever.” ChatGPT’s amazing power has lead to concerns of ChatGPT taking away jobs, helping bad actors spread misinformation, and raising a generation of students who have the attitude that if ChatGPT can write my essay, why should I?

Princeton Senior Edward Tian researched how to detect AI-generated text while working at Princeton’s Natural Language Processing Lab. He writes:

I think we’re absolutely at an inflection point. This technology is incredible. I do believe it’s the future. But, at the same time, it’s like we’re opening Pandora’s Box. And we need safeguards to adopt it responsibly.

Over the Winter holiday, Tian put GPTZero on a free platform. The application uses ChatGPT to determine the level of AI involvement in creating a given text. The system gives each document a “generative AI” score and highlights those sentences deemed most likely to be artificially created. It also provides an API. It was so popular that after three days the platform crashed, a problem that has since been fixed.

Tian used AI-code writing application GitHub CoPilot to help him build the application.

OpenAI has just come out with a similar offering, as Sarah Needelman over at the Wall Street Journal tells us in ChatGPT Creator Releases Tool to Detect AI-Generated Text, Calls It ‘Unreliable’. OpenAI writes that their AI-detector correctly identified 26% of AI-written text as “likely AI-written,” which is not a huge hit rate. On the other hand, it correctly rejects human-written text as AI-generated 91% of the time, which is pretty good. No word on how this compares to GPTZero.

OpenAI warns that the “Is it AI?” Classifier should not be used as the only decision-making tool: it does poorly on highly predictable text, it’s unreliable on text of less than one-thousand characters, does not detect AI-generated computer code, and only works in English. They also note that AI-generated text can be edited to avoid detection. In the final section, “Impact on educators and call for input,” OpenAI asks for input from educators and others who are who are impacted by these issues.