Why AI Still Can't Write Well and Which Half of That Problem Is Actually Yours

Short summary

The real reason AI writing underperforms isn't speculative decoding (a lossless inference optimization) but RLHF training causing mode collapse—models narrow toward bland, safe outputs. Nonfiction writing faces a solvable engineering problem: conflicting objectives between friendly chat and precise technical writing, but fiction requires training datasets that don't yet exist. The author's practical solution: calibrated editing checklists to catch AI writing patterns, which beats waiting for model retrains.

• •Mode collapse from RLHF training, not speculative decoding, is why AI writing sounds bland and cautious
• •Nonfiction writing has a solvable engineering problem; fiction needs datasets and business incentives that don't exist yet
• •Practical fix today: editing checklists calibrated to AI writing patterns, not better models

Generated with AI, which can make mistakes.