I wonder if one of the issues is, LLMs treat all data sources equally, or they don’t really weight the reputation properly (pure speculation, based only on seeing the results). I know that a large portion of code out there, is not written by seasoned experts, so rather naive code is the fodder for AI. It often gives me stuff that works great, but is rather “wordy,” or not very idiomatic.

For example, court cases mentioned in fictional accounts. If they are treated as valid, then that could explain some of the hallucinations. I wonder if SCP messes up LLMs. Some of that stuff is quite realistic.

I also suspect that this is a problem that will get solved.

I assume you mean this: https://en.wikipedia.org/wiki/SCP_Foundation ??