A security researcher who goes by lcamtuf has published a deeply unsettling account of what happens when AI-generated children's books go wrong—or rather, when they go horribly, horribly right in all the wrong ways. After compiling roughly 220 AI-generated children's titles on Amazon, the researcher decided to actually purchase one: specifically, an Amazon #1 category bestseller in children's encyclopedias. What arrived was not education. It was body horror. The piece explains why AI operators target children's encyclopedias with such ferocity. First, they probably sell well—most kids in developed countries receive one at some point. Second, the buyer isn't the reader; books are judged by their covers and purchased as gifts by relatives who won't inspect the interior closely. Third, unlike fiction, these reference works don't risk infringing on any closely-guarded intellectual property, making it easy for AI-generated knockoffs to undercut traditional publishers without legal exposure. The irony cuts deep here. We're told that frontier models surpassed PhD-level intelligence in summer 2025. Most of these books were published mid-2026 and the artwork appears to come from a flagship model belonging to a major US-based lab. By all accounts, we should be well past the point where AI can handle simple children's encyclopedia entries. But apparently, 'well past' includes generating images of a child with their face peeling back to reveal another face underneath. Other illustrations documented in the piece include: reflections that reach out to grab the viewer, children waking up on alien planets screaming for help while no one believes them, cats that have returned from somewhere wrong, and nature scenes where beasts and trees fuse into malevolent pulsating masses with vines wrapping around ankles as disembodied voices whisper. All of this comes from a #1 category bestseller. The researcher notes that rankings and reviews can be faked, but browsing relevant Amazon product categories reveals this content is widespread—suggesting real sales to real families.

The Core Problem

The researcher's conclusion is stark: 'It's entirely possible that the models of tomorrow will be able to generate flawless children's encyclopedias. But until then, we're messing up some kids.' This isn't a story about AI being bad at generating text—the sameness problem has been covered elsewhere. It's about what happens when the technology is deployed in production environments, trusted by parents who assume Amazon's bestseller algorithms mean something, and potentially shaping the mental development of young children with nightmare fuel disguised as educational material.

Key Takeaways

  • Children's encyclopedias are a prime target for AI-generated content due to reliable sales, non-reader buyers, and no IP infringement risk
  • The #1 category bestseller on Amazon contained deeply disturbing body horror imagery alongside educational text
  • Current flagship models from major US labs are producing this content in mid-2026 despite claims of PhD-level intelligence capabilities
  • These books shape young minds while operating largely under parental radar due to gift-giving purchase patterns

The Bottom Line

This isn't a future problem. It's happening now, on the world's largest e-commerce platform, with products rated highly enough to stay visible. Someone needs to audit what Amazon is actually selling to children—and parents need to stop trusting algorithmic bestsellers as quality signals for anything that requires actual editorial judgment.