The developer warned that the text classifier can be easily fooled.
OpenAI, developer of ChatGPT, released a new tool that predicts whether a text was written automatically, although it warns that it is not entirely reliable. The multiple possible uses of OpenAI’s artificial intelligence sparked debate around the intention of some of them, as is the case of the plagiarism.
The tool developed by OpenAI is revolutionizing technology due to its ability to write texts that seem to have been written by a person, making it almost impossible to detect if it is the work of artificial intelligence.
ChatGPT can generate meaningful texts on almost any topic, which led OpenAI to develop a tool capable of detecting whether or not a text was written by its artificial intelligence.
The result is now ready and consists of a tool that is born from a modification of GPT, which is the base technology used by OpenAI for the development of its popular bot. The developer named it ‘AI text classifier’ and clarifies that its purpose is “to predict the probability that a text has still been generated by AI from various sources.”
The new tool filters the text and returns a result on a scale of five possibilities, ranging from “very unlikely” to “most likely generated by artificial intelligence.” The developer specified that for a correct analysis of the text, a minimum of 1,000 characters must be provided, between 150 and 200 words.
OpenAI warns that the tool “is not always accurate, it can mislabel both AI-generated and human-written text.” The company explained that the text classifier can be easily fooled by adding fragments written by a person.
The algorithm of the new tool was trained with databases of texts written by adults in English, so “it is probable that the classifier is wrong in texts written by children and in non-english texts“, they explain. With the classifier, OpenAI intends to “promote the debate on the distinction between content written by humans and content generated by AI.”
How the tool was trained
The developer insists that the results can help, “but they should not be the only test“, as “the model was trained on human-written text from a variety of sources, which may not be representative of all types of human-written text.”
The biggest problems were detected in the academic field. Many students used the openai artificial intelligence to plagiarize texts. As a consequence, some Australian and American educational institutions decided to ban its use.
The developer made it clear that the classifier was not trained to detect plagiarism in an academic environment, so it is not effective for this purpose. At OpenAI they are aware that one of the main uses that they will want to give to the tool is precisely to check if a text is written by a machine or by a person.
However, “we caution that the model has not been thoroughly tested on many of the main intended targets, such as student papers, automated disinformation campaigns, or chat transcripts.” In addition, they add that “classifiers based on neural networks are poorly calibrated outside of their data from training“.
With information from La Vanguardia .