The AI market has hit a paradox: the more neural networks generate content, the faster the supply of "organic" data for training them runs dry. We are witnessing labs resorting to radical measures—from buying up and destroying millions of printed books to licensing human faces for synthetic series. This is not just a trend but a systemic crisis reshaping the data economy.
Model Collapse: The Curse of Synthetics
The key threat to modern LLMs is the so-called "model collapse." This term, formalized by a group of researchers led by Ilya Shumailov in 2023, describes the degradation of AI trained on data created by other neural networks. With each cycle, the model loses touch with reality, responses become homogenized, and errors accumulate. The scale is alarming: according to web analytics firm Ahrefs, over 74% of new online content already shows signs of auto-generation, while Epoch AI predicts the depletion of usable training text by 2026-2032.
This explains the frenzy around books published before 2022. They are the last bastion of "purity," guaranteed free of machine-generated noise and deliberately injected "poisoned" data. That is why ISBNdb, a metadata aggregator for 113 million publications, offered AI companies a service for bulk purchasing books in batches of up to a million copies. The scheme is simple: buy, digitize, destroy. After media exposure, the company hastily removed the promo page, claiming it was merely a "test of market demand." However, independent booksellers confirm an anomalous surge in purchases since April 2026, when everything with an ISBN was bought up indiscriminately, regardless of genre or value.
Faces as Commodities: Renting Appearance
A similar situation is unfolding in the visual content market. Courts in China are toughening penalties for deepfakes, while the AI minidrama industry, where over 95% of 128,000 series are created by neural networks, demands "live" faces. Platforms like ActID and New Claw turn human appearance into a tradable asset, paying between $15 and $700 for a license to use an image. Producers pick faces from a catalog: "girl next door," "rugged," "supermodel."
Notable precedents involve British actors whose digital clones, created for training, ended up in propaganda videos by Venezuelan and Burkinabé entities. This proves that once rights are transferred, control over one's own image is lost irrevocably. The Equity union is already pushing for legislative restrictions, but the market moves faster than regulators.
The Economics of AI Raw Materials
The price gap is striking. A book on the secondary market costs $2-5, while a model trained on millions of such copies is valued in the billions. The same logic applies to voices and movements: Tesla pays operators up to $48 per hour for recording routine actions to train Optimus robots, and SAG-AFTRA is expanding protections against synthetic voice copies.
We are witnessing a fundamental shift: data that used to be taken for free is now being purchased. But the essence of the deal remains unchanged—the asset irrevocably leaves the owner's control, only now with a receipt. This creates a dangerous precedent where human experience and creativity become disposable material for machines, while legal frameworks lag behind technology.
My conclusion: the AI industry is entering an era of "resource scarcity," and this will have serious social consequences. The data licensing market is only a temporary solution. Long-term sustainability will require either a breakthrough in synthetic training or a new ethic where data owners receive not a one-time payment but a share in the final product. For now, we see a classic colonial model: raw materials are taken, products are sold, and locals are left with empty shelves and lost faces.