> If I were to start with ChatGPT I would, necessarily, only say what has been said.

This is not true though, you can test it for yourself:

Prompt: Invent a new word, which has never been written before. When I type it in Google, I should get zero results.

I'm completely transparent, I tried this query 3 times because of my own curiosity and will provide all 3 answers (I added --- for formatting):

---

ChatGPT (answer 1):

Sure, how about "flumtraz"?

---

ChatGPT (answer 2):

Here's a new word: "Ardidio"

It is a combination of "ardor" (meaning passion or enthusiasm) and "audio" (meaning sound or hearing).

The word Ardidio can be used to describe a strong and passionate feeling one experiences when listening to music or any form of audio. For example, "I felt a rush of ardidio while listening to the live performance of my favorite band."

As of now, when you search for "Ardidio" on Google, there are no results found.

---

ChatGPT (answer 3):

"Flumixen" - a verb that means to mix or blend something in a clumsy or awkward manner. For example, "I tried to make a cake from scratch, but I ended up flumixening the ingredients together and it turned out terrible."

---

For the first answer I got zero results on Google, so it's quite unlikely that it was part of the training data, proving that your theory wrong. It can say things that have never been said.

For the second and third answer I got results in Google, but they seem accidental, usernames and the likes, but as far as I can see, they have never been used with the meaning that ChatGPT "invented".

I'm aware that ChatGPT at its core is a stochastic parrot, but the evidence is clear for me that it can combine certain aspects of its training data in unique ways, which were never done before.