Stanford University develops 'DetectGPT' to detect sentences generated by ChatGPT



'ChatGPT' is an interactive chat AI that can output highly accurate sentences, and its accuracy is at a level where it is difficult for humans to distinguish between sentences written by humans and sentences written by ChatGPT. However, since the text output by chat AI such as ChatGPT has characteristics unique to AI, Stanford University is developing ` ` DetectGPT '' to detect sentences created by chat AI like ChatGPT.

Detect GPT

https://ericmitchell.ai/detectgpt/

Stanford introduces DetectGPT to help educators fight back against ChatGPT generated papers - CDJapan
https://www.neowin.net/news/stanford-introduces-detectgpt-to-help-educators-fight-back-against-chatgpt-generated-papers/

The academic world is suffering from dealing with ChatGPT, which has high accuracy but also has the drawback of writing meaningless content in a way that looks like it, and the academic journal Science has revised its policy not to recognize chat AI as an author are being implemented .



On the other hand, it is known that Turnitin, which provides a service to detect plagiarism, plagiarism, and copy and paste in papers, is developing a tool to detect text written in ChatGPT.

Plagiarism / plagiarism detection service Turnitin is developing a text detection tool written in ChatGPT - GIGAZINE



`` DetectGPT '' developed by Eric Mitchell and others at Stanford University is a chat AI-generated text detection tool similar to that developed by Turnitin.

Since text sampled from a large language model tends to occupy the negative curvature region of the model's logarithmic probability function , DetectGPT takes advantage of this to determine if a sentence was produced from a particular large language model. Defines a new curvature-based criterion for determining whether

There is no need to train separate classifiers, collect datasets of real or generated sentences, or explicitly watermark the generated texts, instead combining the log-probability computed by the model of interest with another prior Only random perturbations by passages of the language model learned in .

In the test, it is said that it exhibits superior discrimination power than other detection methods, and the detection of fake news articles generated by the natural language processing AI model GPT-NeoX-20B is 0.81 AUROC of the baseline has been reported to have improved to 0.95 AUROC from

Code and data will be released soon.

in Note, Posted by logc_nt