Update after more discussion: the framing was incomplete.
The real gap isn't generation. Humans have two functions: they generate stray thoughts, AND they dismiss them. A stray thought pops up, gets tagged irrelevant, dropped before it touches behavior. Most mental noise never reaches output because there's a filter.
LLMs have no filter. Anything in the context window is treated as valid input. You can't un-see context. So a thought injector without a dismissal function doesn't make an LLM more human, it just corrupts what it's doing.
Also: temperature is not the same as widening the semantic radius. Temperature flattens the token probability distribution (weird phrasing, still on-topic). Widening the radius changes WHICH concepts get considered at all. Neither one is the dismissal function.
Better question: can we build the dismissal function first? Give an LLM the ability to generate many associations and throw most away. Without that, more thoughts just means more noise.
