Introducing GPTZero, a revolutionary new tool developed by Edward Tian, a 22-year-old computer science professor at Princeton University. GPTZero is designed to detect fraud and plagiarism in text generated by ChatGPT, an AI-based chatbot. GPTZero uses two indicators – confusion and explosiveness – to identify whether the text was written by a human or an AI algorithm. Confusion measures the complexity of the text, while explosiveness measures sentence variation. The tool has already attracted more than 7 million hits and was used by more than 30,000 people, prompting the free platform to provide additional server resources to better handle increased internet traffic. With GPTZero, we can now combat AI-based plagiarism and bring transparency to the black box technology that AI is based on.
Cutting corners: People are responding to potential misuse of ChatGPT by developing new tools to detect fraud and plagiarism. A student has developed a tool that can potentially (and quickly) determine when a piece of text was created by an AI rather than a human typing.
ChatGPT can write code or school essays and any content in response to a human (text) prompt. The OpenAI chatbot is also easy to abuse, creating new content to cheat on exams or fill entire websites with useless but outwardly convincing textual garbage. In response, Edward Tian, a 22-year-old computer science professor at Princeton University, created a tool called GPTZero.
Tian spent his winter vacation developing a platform that can determine who created this essay: a human or ChatGPT algorithm. GPTZero, which is still in beta testing, uses two different indicators, “confusion” and “explosion,” to identify human-generated or AI-generated text passages. Tian says we “deserve to know” when a piece of text was created by a self-aware brain or a computer algorithm.
Confusion measures the complexity of the text, or rather how “puzzled” GPTZero is by the text: a higher level of confusion implies a high level of text complexity, meaning that the passage was probably written by a human. When confusion is low, GPTZero is likely to find text patterns that it is more familiar with, meaning that the text is most likely AI-generated.
here’s a quick demo with John McPhee’s “frame of reference”. pic.twitter.com/WphxfxxFdr
— Edward Tian (@edward_the6) January 3, 2023
Explosiveness, on the other hand, measures sentence variation. People tend to write faster by using longer or more complex sentences alongside shorter ones. AI sentences tend to be more alike, which is a red alert for possibly generated text. Tian admits that GPTZero is not perfect and not reliable, but the system can indeed provide a quick plagiarism test result.
The tool is at least somewhat effective and quite interesting for many professionals working with text: after launching online on January 2, GPTZero attracted more than 7 million hits and was used by more than 30,000 people. It was so popular that the service crashed, prompting the free platform that currently hosts the code to support Tian with additional server resources to better handle increased Internet traffic.
Tian explained Twitter that it wants to do something to curb the growing AI-based plagiarism by bringing some light and transparency to the black box technology that AI is based on. “We really don’t know what’s going on inside” ChatGPT’s algorithms, Tian said, and GPTZero is designed to combat that approach.
Related posts:
