If AI companies avoid AI slop, shouldn't we?
AI companies expect us to consume AI-generated content, even though they themselves vastly prefer human-generated content for training AI models.
SILICON VALLEY, CALIFORNIA (JULY 23, 2026) — 404 Media reported recently that AI companies are buying pre-AI books for training AI.
One major source for book data is a company called ISBNdb, which started out as a source of book metadata for libraries and book sellers, then pivoted to also offering digitized books for training AI. The company claims to be the world’s largest repository of book data and contains around 111 million books.
ISBNdb says that books printed before 2022 largely don’t contain AI-generated content, and are therefore coveted for AI training.
AI companies are also especially keen on Reddit data. One of the reasons for that is that much of that content is generated by people, rather than AI. (They also like it because the topics are many, the information is always being updated and the voting system provides the bonus of human-judgment on the quality of the information.)
While AI companies are flooding the world with AI-generated content, they themselves try to avoid it. The main worry is a fear of “model collapse,” which is what happens when AI is trained on AI data.
“Model collapse” is the process of declining accuracy and quality of output as AI models are trained on multiple generations of AI-generated content. With AI trained on AI content, rare or unusual ideas get lost, the output becomes more generic and average, and hallucinations get validated as true, leading to their elevation in the possible answers chatbots give.
READ: The Attachment Economy, Computerworld, Gastronomad newsletter, Gastronomad book
LISTEN: Superintelligent, TWiT, Gastronomad podcast
JOIN: The Gastronomad Experience
FOLLOW: Gastronomad on Surf Social, Machine Society on Surf Social, Bluesky, Reddit, Notes, Mastodon, Threads, X, Instagram, Facebook, Linkedin!
LOOK AT: Mike Elgan Photography
Here’s the problem
But what’s the difference between AI training data generated by humans who get their information from AI? “Model collapse” still happens — in fact, it’s already happening — but with people plugged into the process.
When a generation of content by writers using AI gets fed back into the system to train AI, then that AI training is pushed back out to writers, who write based on AI-influenced AI and churn out even more homogenized works, and then that content is then fed back into the system, you can see how “model collapse” works even when people are involved.
One study looked at part of the phenomenon. They found that AI-generated story ideas improved the creativity of individual writers, but made the ideas converge with those of other writers using AI to generate ideas, homogenizing the totality of stories (just like in “model collapse”). This study and this also came to similar conclusions.
These studies focused on decreased creativity and diversity of opinion and ideas. But they didn’t address human mediated AI “model collapse.”
It’s nice to see what companies like Substack are doing. (Note: This newsletter is published on Substack.)
