A writer reflects on why good writing remains hard for AI, drawing on 16 years of experience with human editors and more than 10 AI writing systems.

The author has fine-tuned models to mimic a personal voice, kept embarrassing AI-assisted drafts public as proof, and tested multi-model editorial councils modeled on code review.

The experiment reveals fundamental obstacles: each AI has a stubborn voice identity—Gemini beams sunshine, Claude is languid but sharp, OpenAI Codex is dispassionate—and when asked to edit collaboratively, the models produce inconsistent, tone-deaf feedback.

One model described a draft as “500 words of intelligent observation without a single indelible sentence” and suggested picking a single angle.

The author concludes that AI’s ability to synthesize images, video, and text makes anything creatable, but authenticity comes from imperfection: the pops of vinyl, solar flare on Kodachrome, imperfect analogies, and punctuation quirks.

AI may eventually parrot authentic writing, but the imperfections are what make writing good.