the detection method is basically to strip a textual corpus of stopwords like a, an, the, but, or, if, then, else, when, etc... then bind the remaining into bi-grams and/or tri-grams without deduplicating any copies of n-grams, and then shuffle it randomly.
If you ask an AI to "reconstruct" the jumbled bag of n-grams and it converges to basically the same message without any additional prompting... it's highly likely AI generated... as this implies the original was speaking in a diction that is preferred by the model's training distribution to be the most likely continuations of any given idea.
Feel free to use my tool if you don't believe me:
https://ngramextractor.netlify.app/
