Researchers Create 'DetectGPT' To Determine Which Papers Are Written By ChatGPT

When OpenAI's ChatGPT began making waves for its content generation capabilities, instances of students using the tool to write papers skyrocketed in the weeks to follow. Now, Stanford has proposed "DetectGPT" to help educators fight back against papers written by the popular AI tool ChatGPT.

Pexels

Countering ChatGPT

This makes DetectGPT among the first tools intended to tackle automated content at educational institutions. The method dictating DetectGPT is based on the idea that text generated large language models (LLMs) like ChatGPT lurk in specific areas "of the negative curvature regions of the model's log probability function," Neowi reported.

Pexels

Using this insight, the team developed a new barometer to assess whether any given sample of text is generated by a machine that doesn't depend on training on AI tool or collecting large sets of data to compare the text against. This might imply that text written by humans falls on the positive curvate regions, but this implication is not clear.

Also read: ChatGPT Clears Law School Exam In US, Writes Essays On Taxation And Constitution

The method is called "zero-shot" and allows DetectGPT to find text written by machines without any information about the AI that might have been used to create it.

DetectGPT was tested on a dataset of fake news articles and the researchers found that it was able to outperform other zero-shot methods that are designed to detect AI-generated text.

Wikipedia

This makes DetectGPT an ideal tool to check for machine-generated text without requiring any data or training. The importance of machine-generated text will also lead to growing interest in systems designed to detect such text.

Also read: Why Meta's Artificial Intelligence Head Thinks ChatGPT Isn't 'Particularly Innovative'

What do you think about DetectGPT? Let us know in the comments below. For more in the world of technology and science, keep reading Indiatimes.com.