Why AI Can't Write Serious Literature: Taste Variance vs Artistic Intent

On September 22, phl43 and teortaxesTex debated "why AI still can't write high-quality literary fiction": whether post-training simply hasn't prioritized the task, or whether deeper structural obstacles exist.

Confirmed

• phl43 posed the question and offered a guess: human literary taste varies so widely that RL may struggle to optimize for it; he explicitly rejects the claim that "AI lacks human genius, so it can't make successful art."
• teortaxesTex noted targeted attempts have been made, but they weren't a priority, and the task itself is extremely hard: literary writing lacks ground truth — unlike coding with tests, instant computational verification, or Lean-style formalization — so online scoring can only rely on language models as judges, whose reliability is questionable.
• teortaxesTex observed that structured training signals generalize: code-corpus pretraining made models better at nearly all tasks; R1, after RL on math and code, writes far better than V3.
• His core argument: the crux isn't "style" — models are excellent style mimics and can even produce kitsch — but the lack of artistic intent.
• phl43 added: even the latest models remain imperfect on deep conceptual questions yet still perform far better than at literary writing; RL on other objectives brings some generalization gains, but doing similar RL for literary creation is both harder and more expensive.

Why it matters

• The conversation pulls "why AI writes bad literature" out of mystical attributions (no genius, no soul) and back to discussable mechanisms: missing reward signals and wide taste variance make the optimization target hard to define.
• The distinction — "style doesn't require intelligence, but artistic intent is missing" — offers a framework for understanding LLM capability boundaries: the imitable parts are already covered by RL generalization, while the non-imitable parts (intent, aesthetic judgment) remain the bottleneck.
• The practical consensus (rubrics, exemplars, step-by-step guidance) also aligns with their experience on agent reliability, offering direct reference value for creators actually using AI.

2026-09-22 ~ 2026-09-22 · 6 related posts

• [source] Why Is AI Still Bad at Literary Fiction? Taste Variance May Defeat RL — phl43 · 2026-09-22
• [source] Why RL Struggles With Fiction: No Ground Truth, Unlike Code, Math and Lean — teortaxesTex · 2026-09-22
• Making AI Agents Work: Rubrics, Examples, and Step-by-Step Bootstrapping — teortaxesTex · 2026-09-22
• Why RL generalizes to reasoning but not literary writing, per AI researchers — phl43 · 2026-09-22
• Logic Training Generalizes Reasoning, But Style Doesn't Seem g-Loaded — teortaxesTex · 2026-09-22
• [source] LLMs lack artistic intent, and writing may resist outcome-based RL — teortaxesTex · 2026-09-22