ChatGPT, a conversational artificial intelligence article generation system developed by OpenAI, provides inspiration for characters to write, but it also brings shocks to the academic community, worrying that someone will use AI to do homework and write papers. OpenAI just launched a free identification tool to try to help people identify whether the article was generated by AI.
This tool called “AI Text Classifier” is very simple to use. Just paste the article into the input field of the webpage and press the “Submit” button, and the analysis result will be given in a few seconds, but the article must be at least 1,000 bytes (about 150 bytes) -250 English words). The results of the analysis are divided into 5 levels: “very unlikely”, “unlikely”, “unclear if it is”, “possibly” and “likely” . Like ChatGPT, this resolver also requires an OpenAI account, and it also needs to use a VPN to go to a foreign country to log in and use it.
AI Text Classifier:click here
The editorial department of “PCM” tried to use ChatGPT to generate a paper on distinguishing AI-generated articles, and handed it back to the distinguisher for recognition, and obtained a “possibly” result. Another news report copied from the Internet was judged to be “unlikely”, which shows that it has a certain degree of accuracy. However, the Chinese articles generated by AI usually cannot meet the resolution requirements due to the small number of words generated.
However, OpenAI also stated that it is impossible for them to reliably detect all AI-generated articles. They pointed out that this resolver has several limitations, including low resolution when the number of words is less than 1,000 bytes, and articles written by children or non-English language speakers are also easily misjudged as AI-generated articles. This tool is also mainly in English For training; some words that are easy to guess, such as the list of the first 1,000 prime numbers, are also impossible to distinguish whether they are made by humans or AI because the results must be the same. In addition, humans can easily modify the content of AI-generated articles to evade classifiers.
OpenAI admits that the resolution of this discriminator is not high, with a positive and positive resolution of 26% for English articles, and a false positive rate of 9%. However, they still make this tool public, hoping to provide auxiliary tools for the academic community troubled by AI-generated articles, and to collect user opinions to continuously improve the classifier.