After Stanford University announced that it has a solution to detect texts written by AI language models such as ChatGPT, OpenAI, the company that created that AI, has also released a similar tool. However, it seems that this is an early version, which doesn’t work very well at first, despite the fact that it is supposed to detect texts written based on its algorithms.
AI creator tries to provide solutions for text detection with ChatGPT
There is now a new tool called “AI Text Classifier” on the OpenAI website. It allows users to enter text into a box, which is then parsed and classified. To work correctly, the classifier requires text of 1,000 characters, or about 150-250 words. It also currently only works in English, as it has been trained on such texts so far.
Analyzer results are categorized as “very unlikely”, “unlikely”, “unclear”, “possible” and “likely”, relative to whether or not they are written by AI. Thus, if you get on a parsed text the classification likely, most likely, that text was written by AI, while “very unlikely” should be a text 100% written by a real person.
OpenAI however says that it has currently achieved a 26% correct rate for texts written by AI, while it has classified texts written by humans as written by AI 9% of the time. Of course, this classification process will be continuously developed and can be improved as more people use it and “train” it to find templates that AI uses in composing texts.
The company also warns those who will want to use AI Text Classifier to check texts. This should only be a starting point for a plagiarism investigation, and a verdict should not be given based on this analysis alone. It also mentions again to only use it for English. Furthermore, it is not good for use in marking tests whose answers are unique, and not interpretable. An example being the stringing of the first 1,000 prime numbers, which are always the same, so both an AI and a human could give an identical answer.