Frontier AI models supposedly surpassed PhD-level performance by summer 2025, according to their creators. Yet illustrated children’s books produced with those same models are sitting at the top of Amazon bestseller charts with illustrations that lean far more toward body horror than educational content. This is the documented finding of Michal Zalewski, aka lcamtuf, a security researcher well known in the technical community.
AI-Generated Children’s Books Topping Amazon Charts
Zalewski published a piece on his Substack in late June 2026 after buying a bestseller from Amazon’s children’s encyclopedia category. The book was AI-produced, illustrated with what looks like the output of a diffusion model from a major American lab. The result: unsettling anatomical distortions, limbs pointing in the wrong directions, faces that seem to melt into the background, animals fusing with trees.
This is not an isolated case. Zalewski had already catalogued around 220 AI-generated children’s books in a previous article. He notes that if you browse the relevant categories on Amazon, these publications are everywhere.
Why Children’s Encyclopedias Are a Prime Target
Zalewski identifies three structural reasons why children’s encyclopedias are a preferred target for low-cost AI production.
- Market volume: virtually every child in developed countries receives an encyclopedia at some point.
- The buyer is not the reader: the book is typically bought by a relative, judged by its cover, without anyone reading the content first.
- No rights to worry about: unlike fiction, a generic encyclopedia carries no risk of plagiarizing a protected work. Content can be generated at scale with no legal friction.
These three factors combined create a strong economic incentive to flood this segment with automated content, regardless of its actual quality.
The Gap Between Lab Rhetoric and Deployment Reality
This is the core irony Zalewski is pointing at. AI labs communicate about expert-level performance in complex technical domains. At the same time, their models produce illustrations aimed at five-year-olds that contain anatomical errors visible at a glance.
This gap is not trivial in the context of AI-generated children’s books. Children’s encyclopedias shape the first representations of the real world in young readers’ minds. A poorly proportioned cat or a hand with seven fingers are not just aesthetic imperfections: they pass on a distorted picture of reality to a mind that is still forming.
Zalewski puts it plainly at the end of his article: it is possible that tomorrow’s models will produce perfect encyclopedias. But in the meantime, we are doing damage to children.
A Structured Market, No Guardrails in Sight
The phenomenon is not limited to a handful of amateur publications. Platforms like lullaby.ink were already comparing and ranking fully automated illustrated book generators, text and image alike, as of June 2026. The marketing pitch focuses on personalization and speed of production. The question of visual representation quality does not appear in any of these comparisons.
No response from the AI labs involved or from publishers has been documented at this stage. The precise identity of the models used rests on Zalewski’s inference from visual characteristics, not on any official confirmation.
The engagement around the original article, roughly 100 points on Hacker News on the day of publication, indicates the topic hits a nerve in the technical community. The criticism is not fringe.
Key Takeaways
- AI-generated children’s encyclopedias are occupying top spots in Amazon rankings, with illustrations showing the anatomical distortions characteristic of diffusion models.
- The children’s segment is economically attractive for mass AI production: solid volume, buyers who pay little attention to content quality, and no rights to manage.
- The gap between the performance promises made by labs and the reality of their outputs in such a sensitive context raises a genuine question of responsibility.
- No formal guardrails or responses from industry players have been documented to date.
- This is a story worth following: potential regulations, platform responses, and how editorial practices evolve in the months ahead.
If this kind of drift concerns you from a security or software quality perspective, feel free to discuss it. Follow the blog for upcoming analyses.
Sources
- Michal Zalewski (lcamtuf), “AI children’s books, body horror edition”, Substack, June 26, 2026: https://lcamtuf.substack.com/p/ai-childrens-books-body-horror-edition
- Michal Zalewski (lcamtuf), “The 100,000 whys of AI”, Substack, June 21, 2026: https://lcamtuf.substack.com/