The artificial intelligence industry has hit a paradox: the more models are trained on synthetic data, the faster the quality of the systems themselves degrades. In the search for "clean" material—texts created exclusively by humans—labs are ready to take extraordinary measures. This is no longer just about scraping the internet, but about the industrial purchase of printed books followed by their destruction.

Model Collapse: When AI Eats Itself

The term "model collapse" describes the degradation process of a neural network trained on data generated by another neural network. With each cycle, responses become increasingly averaged out and lose touch with reality. This phenomenon was first formalized by a group of researchers led by Ilya Shumailov in May 2023. Their paper "The Curse of Recursion" mathematically proved that regular mixing in of synthetic data leads to the loss of rare events and the accumulation of errors.

The scale of the problem is confirmed by the numbers. An Ahrefs analysis conducted in April 2025 showed that more than 74% of the 900,000 newly indexed pages on Google show signs of auto-generation. And the independent institute Epoch AI predicts that all usable training text on the internet will run out by 2026–2032. Books published before 2022 are becoming a strategic resource—they are one of the few sources guaranteed to be free of "synthetic junk" and deliberate "traps."

Underground Hunt for Tomes

The scheme has already been tested. In August 2024, three American authors filed a class-action lawsuit against Anthropic, accusing the company of copyright infringement. The case materials featured the "Panama" project: the company anonymously bought up print runs of books, cut off the bindings, and scanned them on an industrial scale. In June 2025, Judge William Alsup issued a landmark ruling: digitizing legally purchased books for AI training falls under the fair use doctrine. Now it's enough to keep the receipt—and the origin of the data is documented.

This practice is rapidly being commercialized. The service ISBNdb, which owns metadata for 113 million publications, offered AI companies wholesale brokerage services: buying up batches from 1,000 to 1 million books, guarantees of "purity" from AI slop, and legal protection via NDAs. The promo page directly stated: "The perception problem is real. 'AI company destroyed two million books' is not the headline that will win public sympathy." After being exposed in the press, ISBNdb hastily removed the page, claiming it was just a "test of market demand." However, surveyed booksellers confirmed an anomalous surge in purchases starting in April 2026: one dealer's sales grew from 20 copies to several hundred per week. Buyers were not interested in value or genre—only in the presence of an ISBN.

Face Rental: The New Economy of Appearance

The shortage of "organic" material is not limited to texts. In China, the market for AI micro-dramas is growing rapidly: more than 95% of the 128,000 such films released in the first quarter of 2026 were created using AI. This has generated demand for live, recognizable faces. Platforms ActID and New Claw pay people from $15 to $700 for licensing their images. Users upload photos, and producers select types by categories like "girl next door" or "brutal."

Formally, this looks like a step toward a civilized market. However, lawyer Ile Deng warns: the licensing terms are so vague that it's impossible to understand who ultimately uses the appearance and how. Precedents already exist: British models who sold avatar rights to the startup Synthesia found their digital clones in propaganda videos supporting dictators in Venezuela and Burkina Faso.

Don't Forget the Receipt

The same logic extends to voices and movements. The SAG-AFTRA union has already ratified agreements on protection against synthetic voice clones. Tesla pays up to $48 per hour to people in motion capture suits, and MIT estimates investments in humanoid robotics at $6 billion for 2025.

The scheme is the same everywhere: previously, data was taken without asking; now, people are paid for it. But the essence of the deal does not change—the asset irrevocably leaves the owner's control. Just now with a receipt.

My take: The situation resembles a "gold rush" in reverse—instead of extracting a resource, we are witnessing its total destruction. Legal protection through the fair use doctrine merely legitimizes a barbaric approach. The industry needs sustainable licensing mechanisms, otherwise, within a generation, we will have models trained on themselves—and that will be an irreversible collapse.