I believe it's just complexity and token/compute usage.

You end up chasing diminishing returns as well (100% or even 95% accuracy is just not possible for certain areas of study, especially for niche topics).

It's also 100% unfixable as a premise for the technology. I can enjoy an upscaling algorithm for my retro games to look more detailed at the cost of an odd artifact, but I sure as shit am not taking that risk for information gathering and general study.