Stanford researchers have developed DetectGPT, software that can identify texts written by ChatGPT

Now that ChatGPT is “out in the open” for everyone, all sorts of concerns are starting to emerge about how the general public will use it. Schools and universities in particular are worried that students will bring in homework and projects written by AI texts, without doing the work required to search and write the information themselves, which is part of the learning process. But Stanford University has a solution, DetectGPT an “AI detector” that allows text to be scanned and written passages to be identified using an AI ChatGPT.

ChatGPT will be harder to use to quickly produce papers for school

Stanford researchers say they’ve discovered a way for their software to quickly detect whether text is written by AI. It turns out that there is a template that ChatGPT and other such language models use, which focuses on certain regions in the area of the negative curvature of the logarithmic curve in their likelihood functions. This passage probably sounds a bit like an alien language to many, but such AIs use probabilities to generate content in text or image format.

The DetectGPT solution does not use such “training” AI methods to detect AI-written text. This is a universal solution that can thus identify texts directly, using any text and without having seen such texts in the past. This is why DetectGPT is called a “zero-shot” solution, which seems to work with very high accuracy. Moreover, for this reason, the software developed at Standford not only works for ChatGPT, but also for other similar language models, or various modified versions of OpenAI’s AI.

Certainly, it will be used to check the work of Stanford students, but DetectGPT could also become a very good tool in many other areas, where the actual knowledge of those submitting various papers is very important. Educational admissions, job interviews and other such situations could benefit from an AI detector.