At a recent Day of Learning presentation about zero-cost textbooks, I learned that our economics courses were using LearnVelo, a platform that uses AI to generate a textbook from an instructor’s syllabus and course modules. The college was considering paying the license fee, just as it already supports Pressbooks.
The cost was $20 per student, per course. That is certainly cheaper than a conventional textbook. But I wanted to know what the AI-generated book was actually like.
So I tried it.
The 1,200-page answer
I gave LearnVelo the materials for several of my courses, including Introduction to Film and The Films of Alfred Hitchcock. It generated complete textbooks with remarkable speed.
The Hitchcock book was approximately 1,200 pages long.
This is one of the strange things about generative AI: abundance can look like quality from a distance. A 1,200-page book appears comprehensive. It may even feel miraculous. But students do not need every available sentence about Hitchcock. They need the right material, in the right order, at the right moment.
I downloaded the PDFs and fed them to two other AI clients. I asked those systems to evaluate the books for readability, accuracy, organization, and usefulness to students.
In other words, I asked AI to grade AI’s homework.
Plausible is not the same as usable
The books contained a great deal of solid background information. They also contained the familiar fingerprints of AI-generated prose: repetition, generic organization, padded explanations, and the strangely airless voice of a textbook no one has actually taught from.
The problems were not merely stylistic. The Introduction to Film book often introduced terminology as vocabulary to memorize before asking students to notice what a film was doing. The Hitchcock book treated Rope as one continuous shot, called Blackmail simply Britain’s first talkie, and muddled the production history of Notorious.
Those are fixable errors. The more important problem was that the books knew the subjects, but they did not know my courses.
They did not know why I teach one idea before another, where students tend to get lost, which examples open up a room, or when a term becomes useful because students have already seen the thing it describes. A syllabus gave the AI a table of contents. It did not give it thirty years of teaching judgment.
The books knew the subjects. They did not know my courses.
So I kept the machine and changed the process
The experiment was not a failure. In fact, it clarified exactly what I wanted.
I wanted the speed and reach of AI, but not a sealed product created by someone else’s system. I wanted to control the structure, the emphasis, the examples, the corrections, and the voice. I also wanted the finished books to be genuinely free, not inexpensive only because someone else was paying the per-student fee.
I moved the project to Pressbooks and used my preferred AI clients to rebuild the textbooks according to my own specifications. We shortened chapters, reorganized the books around my actual modules, corrected factual problems, replaced repetitive recall questions with activities and interpretive problems, and made the material more accessible for online reading.
The process remained thoroughly AI-assisted. AI helped gather established background information, propose structures, revise prose, flag questionable claims, and convert unwieldy documents into usable course materials. But I made the decisions about what belonged, what was accurate, what students needed, and what the books should sound like.
What AI changed
It would be easy to describe this as AI saving me the work of writing textbooks. That is not quite what happened.
AI saved me from having to produce every first sentence from a blank page. In exchange, it gave me a different kind of work: selecting, checking, questioning, restructuring, rewriting, and occasionally asking, “Where on earth did you get that?”
That is not the disappearance of scholarship. It is scholarship performed at a different stage of the process.
The finished books are not products I could have obtained by pressing a button. They are course-specific, revisable, and free to my students. They are also unfinished in the useful sense. If an error appears, I can correct it. If scholarship changes, I can update them. If students show me that an explanation does not work, I can write a better one.
For me, that is the real promise of AI in education. Not instant authority. Not infinite content. Not replacing the instructor.
It is using the machine to gather and shape what is already known, while keeping human judgment exactly where it belongs: in charge.