Mocked UCLA AI Textbook Reached Class—But Evidence Stops at Self-Reports

UCLA’s disputed AI-assisted literature textbook moved beyond its mocked debut and into Comparative Literature 2BW during 2025. A February 2026 post-course interview records professor Zrinka Stahuljak’s account of broader classroom participation, increased office-hour attendance and more time for discussing primary texts.

What remains unresolved is whether those changes produced better learning. The available evidence consists of the instructor’s retrospective and individual student observations, without published comparative data on grades, retention, comprehension or other outcomes for the course.

The textbook became part of a real UCLA course

The project began as more than a conventional book with a few automated extras. UCLA’s December 2024 account of Comp Lit 2BW identified it as the Humanities Division’s first course built around Kudu; described a winter 2025 package of textbook, assignments and teaching-assistant resources; set the student price at $25; explained that the digital book could be printed, used with audio readers and updated during the term; and presented the cover’s nonsense lettering as an intentional prompt about language, history and fiction.

Calling the package “AI-generated” can obscure how much human control its production involved. Stahuljak supplied notes, presentations and videos from earlier versions of her survey of literature from the Middle Ages through the 17th century. She and a student reviewer vetted the resulting material, while the platform’s question-answering feature was limited to resources approved for the class.

The resulting workflow was therefore closer to machine-assisted course publishing than to an autonomous chatbot writing and teaching a subject. Faculty-selected material defined the knowledge base, people reviewed the output, and the instructor remained responsible for the course. That distinction matters because any benefit cannot be assigned to Kudu alone.

The cover criticism exposed a genuine credibility problem

The cover paired manuscript-like imagery with malformed lettering that appeared unreadable. Whatever the intended classroom purpose, the image resembled a familiar failure of generative-image systems and became the public’s first evidence of the project’s quality.

The official explanation—that the lettering was deliberately nonsensical—clarifies the designers’ intent but does not erase the reputational damage. A confusing cover is a particularly risky choice for an educational product whose accuracy, authorship and readability are already under scrutiny.

It also does not settle the quality of the chapters. The cover was a generated image containing pseudo-text; the course content came from Stahuljak’s teaching archive and underwent human review. The visible defect cannot prove that the chapters were inaccurate, just as editorial supervision cannot by itself prove that students learned more.

Students did not use every feature in the same way

A February 2025 report from inside the classroom observed about 100 undergraduates discussing Don Quixote and documented their use of comprehension questions, audio or podcast-style material and Kai, the course-specific virtual assistant.

The reactions were mixed and concrete. One student valued listening to course material while walking across campus. Another appreciated the book’s organization and self-check questions but rarely used its audio or assistant features, preferring to read independently before bringing her interpretation to class.

A third student found the assistant useful when an instructor or teaching assistant was not immediately available. These examples establish plausible uses for accessibility, review and low-stakes clarification. They do not show that the whole class preferred the system or that the features improved academic performance.

The strongest update is still the professor’s own assessment

After teaching with the book, Stahuljak described several visible changes: engagement extended beyond the usual front row, more students brought papers to office hours, and some listened to audio versions of chapters while travelling to class or exercising. She also described replacing some background instruction with primary-source discussion.

The teaching assistants’ role changed as well. Instead of trying to cover every assigned source during a discussion section, they could concentrate on one text, run an in-class analysis and writing exercise, and give students immediate feedback. This is a specific account of how the technology altered teaching time, not merely a general claim that AI made the course more efficient.

Those observations remain valuable but limited. They came from the professor who created, edited and adopted the material, and the public follow-up supplied no baseline participation counts, assessment results or comparison group. Increased office-hour attendance may indicate engagement, but it is not a direct measure of comprehension or durable learning.

What the UCLA experiment actually establishes

The defensible conclusion is narrower than either the ridicule or the later success narrative suggests. The textbook survived an embarrassing public introduction, functioned in a humanities classroom and supported uses that some students and the instructor could describe specifically.

The experiment also shows that an “AI textbook” may be a bounded system constructed from faculty-approved material rather than an unrestricted model searching the open web. That design can reduce the range of information available to the assistant, but it does not remove the need to review generated text, instructional choices and student outcomes.

It does not establish that Kudu teaches comparative literature better than a conventional course or that another instructor would obtain the same result. Stahuljak’s archive, editing decisions, classroom practice and teaching assistants were essential parts of the deployment. Until independent outcome evidence appears, the project is best understood as a documented teaching experiment with encouraging self-reported results—not a demonstrated victory for AI-generated textbooks.

Also read:

• 44% of All New Music on Deezer Is AI-Generated — But Almost Nobody Is Actually Listening to It
• The Oscars Banned AI, But AI Probably Didn’t Get the Memo