If you input 300 bits into an LLM, and it outputs an additional 700bits of new information, I’d still rather you tell me those 1000bits than read ann llm’s output of 1000bits.

The primary reason is that human language is becoming a proof of work, that speaking out loud or writing directly indicates that the idea is important enough for a human to express. This is more costly than llm output, which is often just botspam.