"...the unreliability of the LLMs in terms of just making stuff up..."
In other words, it works like the human brain. It's long been known the human brain can very convincingly fill in missing data from a memory by pulling in data from similar memories, or even from beliefs or previously imagined scenarios.