AI still can't write textbooks better than humans, author finds

The author's experience reflects a broader stagnation in AI writing capability that contrasts sharply with breakthroughs in adjacent domains. While models like GPT 5.5 Pro and Claude have become indispensable for mechanical editing tasks — catching typos, refactoring code, and automating document synchronization — their core weakness in long-form composition has not meaningfully improved. The author notes that the most celebrated writing-focused models (OpenAI's GPT 4.5 and Moonshot's Kimi K2) are themselves old, even as models have leapt from "incapable to decent" at search and research tasks and from "okay to superhuman" at coding and mathematics. This divergence suggests that writing excellence is not simply a function of raw model scale or general intelligence, but depends on specific training signals and interventions the field has not yet cracked.

The author's observation about "compounding errors" is particularly telling. Modern models excel at point tasks — finding a single typo, fixing a local bug, generating a code snippet — but when forced to maintain coherence across a chapter or across multiple revisions, they degrade. This is not merely a symptom of insufficient intelligence; the author implies it reflects a fundamental architectural or training mismatch. She contrasts this with how "reinforcement learning from human feedback has been a truly magical solution" for reducing compounding errors in math and code, yet the field has not widely applied equivalent techniques to writing. The author suspects writing lacks "good training data to specifically intervene on it," and that inference-time scaling (spending more compute tokens at generation) and prompt engineering, while helpful, will not unlock multiplicative gains.

For the author's own case, the calculus was pragmatic: she used AI to automate expensive, repetitive labor (syncing documents, formatting, copyediting passes) while reserving the generative core — structuring knowledge, organizing chapters, crafting the narrative spine — for herself. She accepted this asymmetry not because AI has reached parity with human expertise, but because the publishing timeline and field velocity made some time savings justified. Her caution is instructive: she explicitly avoids AI outputs in her blog writing, where voice and point of view are paramount, yet found it acceptable in a reference text where her role as expert gatekeeper could catch and correct AI flaws. This boundary suggests a sustainable middle ground — not that AI writing is solved, but that humans can leverage it productively only when they retain deep domain knowledge and editorial control.