The artificial intelligence 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" human material. And while it used to be about scraping open web pages, the industry has now moved to radical measures—from wholesale destruction of printed books to renting human faces.

The Curse of Recursion

The term "model collapse" was introduced into scientific discourse by an international group of researchers led by Ilia Shumailov back in May 2023. Their work mathematically proved that when trained on data generated by other neural networks, a model loses its ability to correctly reproduce rare events and subtle patterns. Errors accumulate, response diversity declines—and each successive generation of AI becomes increasingly "averaged out" and detached from reality.

The scale of the problem is confirmed by the numbers. According to my analysis of web analytics data, as early as April 2025, more than 74% of the 900,000 indexed pages showed signs of auto-generation. And according to Epoch AI estimates, all the text suitable for training on the internet could run out by 2026-2032. This is why books published before 2022 are gaining strategic value—they are one of the few sources guaranteed to be free of "synthetic junk."

The "Book Under Scanner" Scheme

The market's reaction was swift. ISBNdb, a metadata aggregator for 113 million publications, offered AI companies a specific service: wholesale purchase of books in batches of up to a million copies, followed by digitization and disposal. As it turned out, this is not an isolated case—a precedent has already been set. In June 2025, a U.S. court ruled that digitizing legally purchased printed books for training language models falls under the fair use doctrine. Now it's enough to buy a paper book and keep the receipt.

Significantly, after the investigations were published, ISBNdb rushed to remove the promo page and stated that it was only about "testing market demand." However, independent used-book dealers confirm anomalous purchases since April 2026—sales have grown from twenty copies to several hundred per week. Buyers are not interested in the genre or the value of the publications—only the ISBN matters. Unique copies are irretrievably lost under the scanner.

Faces as Commodities

Similar dynamics are observed in the visual content sphere. Platforms like ActID and New Claw are already building a market for licensing human faces for AI-generated series and advertising, paying from $15 to $700 for image use. In China, where more than 95% of the 128,000 mini-dramas in the first quarter of 2026 were created with AI, this has become a full-fledged industry.

However, as practice shows, control over one's own digital likeness is an illusion. British actors who sold the rights to their avatars found them in propaganda videos of the Venezuelan and Burkinabé regimes. Even official contracts do not protect against unintended use—licensing terms are often so vague that it is impossible to track the end consumer.

My Verdict

We are witnessing the formation of a new economy where human experience—texts, faces, voices, movements—becomes a raw material resource. The price gap is striking: a book for $2-5, destroyed after scanning, helps create a model worth billions of dollars. But the main problem is not the price, but the irreversibility. After transferring rights, a person loses control over their digital footprint forever. For now, the industry is moving along the path of least resistance, but this imbalance will inevitably lead to stricter regulation—the question is only who will manage to profit before that moment arrives.