I wonder... LLMs currently work because they have a huge corpus of data to train against. Data generated by humans. What happens when almost all available data

I wonder... LLMs currently work because they have a huge corpus of data to train against. Data generated by humans. What happens when almost all available data is computer-generated? When no new training datasets are available? Will the models feed back on themselves? Will they stagnate and cease improving? Perhaps we'll enter a world where everything new and unique made by human authors (text, art, source code) is jealously guarded in an attempt to keep the LLMs out - lest even a single model learn it and spread it to all the others almost instantly making the original author's work worthless.