The AI market has hit a paradox: the more models are trained on synthetic data, the faster they degrade. This phenomenon, known as "model collapse," is forcing labs to seek ever more exotic sources of "clean" content—from mass purchases of printed books to renting human faces for deepfakes. I explore why an industry striving for autonomy is now desperately in need of "organic" material.

The Curse of Recursion: How AI Devours Itself

The term "model collapse" was introduced into scientific discourse in May 2023 by a group of researchers led by Ilya Shumailov. It has been mathematically proven: when a model is trained on data generated by another neural network, it loses its ability to correctly reproduce rare events and subtle patterns. With each cycle, outputs become increasingly averaged, and errors accumulate like an avalanche.

The scale of the problem is confirmed by the numbers. By my estimates, over 74% of new content on the internet is already created automatically, and according to Epoch AI forecasts, all publicly available text suitable for training could run out as early as 2032. This creates enormous demand for data guaranteed to be free of "synthetic junk"—and books published before 2022 are becoming the leading candidate.

The Gray Scheme: Books as Consumables

A telling case is that of ISBNdb, a service that offered AI companies wholesale brokerage services for purchasing printed books in batches of up to a million copies, followed by digitization and disposal. After this scheme became public knowledge, the company hastily removed its promotional page, calling it all a "test of market demand." However, independent used-book dealers confirm an anomalous surge in purchases starting in April 2026, when sales grew from twenty to several hundred books per week.

This practice has already received legal approval. In June 2025, Judge William Alsup ruled that digitizing legally acquired books for LLM training falls under the doctrine of fair use. Now it is enough to keep the receipt—and the origin of the data is documented. The economics of this scheme are shocking: a book for $2–5 on the secondary market turns into an asset that helps create models worth billions of dollars.

Faces and Voices: The New Currency

A similar situation is unfolding in the biometric data market. Platforms like ActID and New Claw pay people from $15 to $700 for licensing their likenesses for AI series and advertising. In China, where over 95% of the 128,000 AI micro-dramas released in the first quarter of 2026 were created using neural networks, the demand for "live" faces is enormous.

However, control over the use of this data is extremely fragile. There are known cases where digital clones of British citizens who sold the rights to their avatars were used in propaganda videos of dictatorial regimes. This confirms my long-held belief: after transferring rights, a person effectively loses control over their digital image forever.

The Analyst's Verdict

We are witnessing a fundamental shift: the AI industry, striving for full autonomy, is in fact critically dependent on human "raw material." Whether it is texts, faces, or voices—the pattern is the same: data irreversibly leaves the owner's control, now with a receipt in hand. Until the market develops robust mechanisms for long-term control and fair profit distribution, this asymmetry will only intensify, creating new ethical and legal time bombs.