The book "Your Life in Numbers: Modeling Society Through Data" by Pablo Jensen provides an interesting insight on this, notably the chapter "We are not social atom" [1]. The author argues notably in an interview [2] that for social behaviors, prediction models work hardly better than just saying that next year will be the same as last year.

The main illusion of scientific models for predicting the future is that it worked well for group of simple objects like planets or atoms. However, applying this approach to groups of humans is often disappointing.

[1]: https://link.springer.com/chapter/10.1007/978-3-030-65103-9_...

[2]: https://news.cnrs.fr/articles/why-society-cannot-be-modeled

Yeah. If you can check reproducibility with AI then you can just randomly generate a million of texts and predict what will reproduce. But AI is not this magic since it is trained on past texts.