OpenAI announced on Tuesday a new tool called AI Text Classifier, which is intended to distinguish between AI-generated text and words that were written by a human.
The brand’s latest tool comes after criticism of its ChatGPT text generator. While it has been popular since its November 2022 release, the chatbot has been embroiled in plagiarism concerns, with the tool already being banned from schools. However, reports have also surfaced of students using ChatGPT to create essays for take-home assignments.
Making the AI Text Classifier available is OpenAI’s effort to keep its brand honest. OpenAI said in a blog post that it would like feedback from parents and teachers on the quality of its plagiarism-detection tool. However, it does note that the Classifier has some limitations at the moment, including not working well for texts shorter than 1,000 words, not working well for languages other than English, being easily duped if AI text is edited, and producing false positives, where the tool incorrectly identifies a human-written text as AI, among others.
OpenAI added that the Classifier has an approximate 26% rate of correctly identifying an AI text as likely AI-written (true positives),and a 9% instance of incorrectly identifying a human text as AI-written (false positives).
When testing the AI Text Classifier, I compared the ChatGPT prompt “write a 1,000-word essay on the number of stars in the sky” to the text from my article “The 6 biggest problems with ChatGPT right now.”
Logging into the Classifier is easy if you already have an OpenAI account and servers are not congested. You just paste your text into the text box and select submit. The result will display below. The tool accurately identified the text from my article as human-written and the ChatGPT-written text as AI-written.
There are a number of other plagiarism-detection tools already available, with varying degrees of accuracy. GPT-2 Output Detector is another tool developed by OpenAI. Others include Writer AI, Content at Scale, GPTZero, and Giant Language Model Test Room (GLTR). We tested these tools and similarly found varying degrees of accuracy between the identification of ChatGPT text and human text.
It’s not clear how the AI Text Classifier will be implemented moving forward, but ChatGPT is already solidifying itself in several industries, from journalism to business and technology. With all the proper tools in place, it likely just becomes a matter of ethics.
