I spent months writing a textbook about artificial intelligence. During that process, I used every AI writing tool I could find. The experience left me with a concrete prediction: humans will still write the best textbooks for another two to five years, because right now AI saves only about 10 to 20 percent of the total effort.
That number surprises people. They assume that because language models can generate fluent paragraphs, they must be able to generate books. They can, in the same way that a laser printer can generate a novel. The output exists, but the hard parts of textbook writing happen before and after the drafting stage.
Where AI Actually Helps
The drafting phase is where AI shines. If I need a clear explanation of backpropagation or a summary of transformer architecture, a model can produce a decent first pass in seconds. That is genuinely useful. It removes the blank-page problem and gives me clay to shape rather than nothing.
But a textbook is not a collection of clear explanations. It is a curriculum made physical. Every chapter must build on the last. Every example must use consistent notation. Every exercise must test the specific concepts introduced in that section, at the right difficulty gradient, with solutions that do not accidentally introduce notation from three chapters ahead.
AI is mediocre at all of these constraints. It forgets my notation choices. It generates exercises that assume knowledge I have not taught yet. It writes explanations that are clear but pedagogically backwards, starting with the abstract principle and only later (if ever) grounding it in an example.
Where Humans Still Dominate
The real work of textbook writing is structural. I spent more time on the dependency graph between chapters than on any individual paragraph. I had to decide: does attention mechanism come before or after the full transformer block? Should the ethics chapter sit at the end as a capstone, or early as a frame? These are not writing problems. They are curriculum design problems, and they require knowing what a student will misunderstand before they misunderstand it.
AI cannot yet model the student mind with enough fidelity to anticipate these failure modes. It can generate text that sounds correct to an expert, but experts are not the audience. The audience is someone who has never seen a loss landscape and will panic if you show them one without warning.
The 10 to 20 Percent Window
That is why my estimate is two to five years. AI currently handles the easiest 10 to 20 percent of the work: generating clean prose, summarizing papers, suggesting analogies. The remaining 80 to 90 percent involves pedagogical sequencing, constraint management, and the kind of anticipatory empathy that comes from having been confused by the material yourself.
I suspect that gap will narrow. Multimodal models that can read diagrams, agentic systems that can maintain consistency across hundred-page documents, and fine-tuning on pedagogical corpora will all help. But textbooks are high-stakes artifacts. A wrong explanation in a blog post is embarrassing. A wrong explanation in a textbook shapes thousands of minds. The human gate will stay in place longer than the technology strictly requires, because the cost of being wrong is too high to outsource.