Because “ai” ad we colloquially know today are language models: they train on and can produce language, that’s what they are designed on. Yes, they can produce images and also videos, but they don’t have any form of real knowledge or understanding, they only predict the next word or the next pixel based on their prompt and their vast examples of words and images. You can only talk to them because that’s what they are for.

Feeding research papers will make it spit research-sounding words, which probably will contain some correct information, but at best an llm trained on that would be useful to search through existing research, it would not be able to make new one