Text detector written with ChatGPT!
How to distinguish between texts written by ChatGPT and those written by humans? The detector achieves this 100%.
Mathilde Fontes, editor-in-chief of the scientific journal Epsilon, Today he tells us about an impressive detector, which thwarts the already exceptional performance of ChatGPT: How do we distinguish texts written by AI from texts written by humans?
franceinfo: American researchers have just developed a detector that can distinguish between texts written by humans and those written by artificial intelligence?
Mathilde Fontes: Yes, big issue. We remember the historic strike of Hollywood screenwriters: a month ago they were given guarantees that a minimum number of people would continue to write scripts. But we still need to distinguish these texts from others, and set controls for them…
The problem also arises on the scientific side: most scientific journals have changed their publication guidelines to prohibit the use of AI – because ChatGPT is amazing, incredibly powerful, but it can sometimes “hallucinate”: write things that are completely false, without knowing why, and add… Lies in scientific articles.
This team from the University of Kansas focused specifically on science. She trained a program to study scientific articles in chemistry journals…
Is this AI that has been trained to detect AI?
Yes, AI is controlling AI – it’s a bit amazing. It is an educational program, trained on 100 articles written by humans, and 200 articles written by ChatGPT specifically for this occasion. Researchers have given it features for examining writing style: Difference in sentence length, repeating some words, punctuation marks…
It was able to detect scripts written by ChatGPT, in 98 to 100% of cases. He even succeeded in magazines for which he had not trained.
This is the first GPT cat detector?
No, there were others. OpenAI, the company that designed ChatGPT, has already developed one. Another one is called ZeroGPT. But their performance is less impressive: they found AI in only 10 to 65 percent of cases.
The strong point of this new detector is that it is specific: it has been trained on a specific type of text: scientific publications in chemistry journals. As a result, it is more accurate. This does not mean that he cannot process other texts if we modify his training.