The tool was introduced a couple of months after OpenAI drew public attention to ChatGPT, a bot that can write texts like a person. Last week, Microsoft, which has already invested heavily in OpenAI, announced a multibillion-dollar investment to integrate the startup’s developments into its services.

Many educational institutions have restricted the use of ChatGPT by students, due to concerns that the bot could harm the quality of education. Sam Altman, head of OpenAI, jokingly said that it was the same with calculators. However, he also said that the company could help teachers recognize “fake” work done by AI.

The development of the new tool is ongoing, but currently it still makes many mistakes. OpenAI is waiting for feedback from teachers and parents. According to OpenAI, the classifier, when evaluating test texts in English, successfully identifies 26% of AI-generated texts as “probably written by AI”, but incorrectly classifies 9% of human-generated texts as written by AI.

This is not the first attempt to create a tool for recognizing machine-generated texts. Recently, Princeton University student Edward Tian introduced the GPTZero to educators. OpenAI previously released its AI-generated text recognition tool in 2019 along with the LLM language model. According to the startup’s employees, the new version of the tool more effectively recognizes texts that are passed off as texts written by a person.

The tool is not effective in parsing texts shorter than 1000 characters, and OpenAI does not recommend using it for languages ​​other than English. Also, AI-generated text can be easily corrected by hand, making it harder to identify. OpenAI already pointed out in 2019 that synthetic text identification is challenging, but the company plans to continue working on system recognition in the future.