MADRID, 30 Jan. (Portaltic/EP) –
Researchers at Stanford University have developed DetectGPT, a tool that is capable of detect if a text has been created by Artificial Intelligence (AI), like ChatGPT, using probability and statistical data.
With the increase in development, fluency, and knowledge that extensive language models are acquiring, students can make use of these tools to carry out written tasks. As the group of researchers who created DetectGPT mentions, “this makes it difficult for teachers to accurately assess if the student has really learned”.
In this sense, one of the AI tools that stands out the most is ChatGPT developed by Open AIsince it presents very effective benefits when it comes to generating coherent texts and with a high writing capacity, which manages to carry out extensive work with just giving a few instructions and a click.
In this framework, it has created DetectGPTthe tool that was created with the aim of recognize and distinguish texts created by AIso that get to detect fraud in the authorship of the texts.
The operation of this tool is based on a statistical method of “probability curvature”, which suggests that the text developed by an AI tends to occupy regions of negative curvature of the log-likelihood function of the model.
Based on this method, the research team developed a new criterion based on this curvature to determine if a text has been generated from an AI or not.
Thus, as the developers explain, DetectGPT only uses the logarithmic odds calculated by this model, and does not require training a classifier or collecting a data set of real or generated tickets.
