Why haven't we have a great AI novel yet?
Because LLMs are compression engines trained on the statistical regularities of human text. When generating stories, they produce a kind of average of all stories in the training distribution. This is why LLM fiction tends to be so generic: competent but undistinguished prose, predictable character arcs etc.
This kind of "average of averages" explains not just why LLM fiction is bad but why it's bad in such a specific and recognisable way. It's never incoherent or offensive. It's never wrong exactly. It's just… beige.
And scaling doesn't work either. More parameters and more data push the model toward a higher-resolution average, not away from averaging itself. You get more polished averages, not more distinctive points. Better beige is still beige.
This is the "ill-defined domain" problem. AI is currently very good in certain domains: code compiles or it doesn't, a proof holds or it fails etc. But literature doesn;t work at all like this. There's no compiler for narrative or readily formalisable criteria for a truly great work of fiction.
So far “telling a satisfying and well-written medium-length story” has proved far harder for LLMs than mathematical proofs, music generation, research reports, code, and many other forms of work.
The technical reasons are pretty clear, but they are supposed to be language models