A random number generator uses a seed value. The larger the seed value domain, the more unpredictable the "randomness" of the generator. The "randomness" of a generator with finite domain(even if it derives new seed values from old values) ultimately becomes predictable and hence useless.

The "AI" systems, as currently designed, could face a similar problem as a random number generator with finite seed value domain. The present AI systems are seeded with information generated by humans or derived from it. But as humans start relying on AI generated information, the seed value domain of the AI systems will become finite as the AIs will start rehashing new information from the same old values or values derived from it. Like a finite seed value domain random generator, the AI systems of today will eventually cease to generate new information. The point of intersection on a graph that plots contribution by AI and humans to the information domain is the one to look out for. At present humans contribute more than AI, but once that graph flips, then the limiting of the information domain starts.

It will be a tragedy if humans lose the ability to generate new information on their own and the AI systems become constrained to a finite information domain. The AI systems will lose their ability to generate new information. It may not happen immediately, but at some time in the future we are going to hit that point.