Journalist of the broadcast ArsTechnica Benj Edwards talked about an experiment conducted during the examination of the accuracy of large language models of artificial intelligence – ChatGPT and the like. The journalist worked with a text detection system generated by AI technologies.

It’s about a tool like GPTZero. At the core of the technology is a large language model trained on large volumes of text, consisting of millions of letter samples and a set of prediction rules that determine whether text is more likely to be produced by humans or artificial intelligence.

However, “exact” systems may well fail to describe complex and multidimensional texts such as the Bible and similar volumes, as well as the US Constitution.

The GPTZero system, while detecting the text of the Constitution, published a definition according to which the US Constitution was written by artificial intelligence. The fault of AI systems is that they have a detection function like “surprise” – how close the raw data produced during analysis is to the material on which the language model is trained.