02/13/2023
ChatGPT, the artificial intelligence (AI) chatbot developed by OpenAI has become quite popular, with people using it for all kinds of things, from answering random questions to getting the chatbot to do their work for them.
Stanford researchers have developed a tool, dubbed as DetectGPT that will help teachers identify content that has been generated using ChatGPT or other similar large language models (LLMs).
Since the introduction of the chatbot, there have been concerns about students getting their papers done using ChatGPT. After New York City public schools banned students from using ChatGPT, its maker OpenAI confirmed that it is working on a tool to detect content generated by ChatGPT.
DetectGPT is a tool that has been designed to detect content generated by ChatGPT and similar tools. According to the researchers, text generated using LLMs such as ChatGPT “occupy negative curvature regions of the model's log probability function”.
Using this, the researchers have defined a new criterion for judging if a passage is generated using an LLM. This allows DetectGPT to use log probabilities to find out if the given passage has been generated using a given LLM. The researchers have claimed that DetectGPT has an accuracy of 95%.
