technology

A student has developed software that detects texts written by ChatGPT

A student has developed software to detect texts written by ChatGPT.

© Getty Images

A computer science student specializing in journalism spent part of his winter break creating GPTZerosoftware that allows you to “quickly and efficiently” determine whether a text is credited to a real human being or ChatGPT artificial intelligence. A new tool that results from the dishonest use made of ChatGPT by some students.

Edward Tian explains that he developed this software with the aim of combating an increase in plagiarism thanks to AI. “There is so much hype around ChatGPT… Is this and that written by AI? We as humans deserve to know!”, writes the young man in a tweet presenting GPTZero. The student also said that many teachers contacted him after his application was published on January 2, telling him of positive results from testing it.

Index of “perplexity”

To determine if a snippet is written by a bot, GPTZero uses two indicators: perplexity and burst. Both are based on dissecting sentences from a text, analyzing whether they consist of complicated words or are constructed in a loosely formatted way. If the perplexity score is high, meaning that GPTZero was “surprised” while reading it, then the text is more likely to have been written by a human. Conversely, a slight perplexity attests that the text could well be the work of an AI.

A first try on GPTZero.

A first try on GPTZero.

© Screenshot Les Numériques

One indicator not being sufficient, Edward Tian has integrated a second one. The burst compares the variations of the sentences. Indeed, humans tend to write with more variation (shape, length, construction), while AI-generated sentences tend to be more uniform.

Tian has published a demonstration video where two texts are compared, one from The New Yorker, the other being a LinkedIn post written by ChatGPT. In both cases, the software managed to distinguish human writing from AI. In his answers to Internet users, the student nevertheless specifies that his software is not infallible, but that it responds to certain basic situations.

Advertising, your content continues below

Related Articles

Back to top button