I agree that AI-written text often has a low content density. I wonder if it's a matter of information theory.

Information theory defines the information of a symbol as being related to how often it occurs and how often it is expected to occur. Something that isn't expected carries more information. (Usually "symbol" is defined as one character or byte, but it could be a word or word part.)

Well, if you think about LLMs that way, they give you the most-probable next word (or word part). That means that they give you less information than normal writing. I suspect that's why it reads as bland, low-content - because it really is low content, in the information theory sense.

Now, it doesn't always give you the most probable next symbol. There is some randomness. And you can increase the randomness by turning up the temperature. But if you do, then I suspect it becomes incoherent more quickly. (Random gibberish may have high information from an information theory standpoint, but humans don't want to read that either.)