DetectGPT – the Tool to Detect Text Generated by AI

It will now be much more difficult to pass off a term paper written by AI as your own. Stanford's developers will equip teachers with the DetectGPT tool, which can accurately determine authorship.

The intensity of the use of large language models (LLM) is growing at a rapid pace today. This process was further accelerated by the well-known chatbot ChatGPT, created based on the GPT-3.5 language model. It turned out that it can write text homework and term papers so well that it is almost impossible to distinguish between the text it generated and the one written by the student. Realizing the scale of the catastrophe, many teachers expressed the expected protest and voiced the need for systems that could determine whether the author of the text is a machine or a person.

Recently, Stanford researchers led by Eric Mitchell presented the DetectGPT detector, one of the first tools to combat AI-generated texts in education. It is based on the ML "zero-shot learning" model (the ability of the system to recognize what it is not yet familiar with). It allowed DetectGPT to accurately identify text written by artificial intelligence. True, without specifying which particular language model was used to create it.

The development team put DetectGPT to the test on a fake news dataset. The results obtained showed that their detector is significantly superior to other methods for detecting text generated by AI. The high performance of DetectGPT suggests that in the future it may well become the most promising method for studying and detecting machine-generated text.