The artificial intelligence training market is undergoing a tectonic shift. We are witnessing LLM development labs entering a new phase—a phase of desperate hunting for "clean" data untouched by synthetic content. This is no longer just a technical problem, but a fundamental challenge to the entire industry.
Model Collapse: When AI Devours Itself
The term "model collapse" describes the degradation of a neural network trained on data generated by other AIs. Each new cycle makes the results more averaged and predictable, erasing the connection to reality. This phenomenon was first formalized by an international group of researchers led by Ilya Shumailov in May 2023, publishing a paper on the "curse of recursion."
The scale of the problem is confirmed by numbers: in April 2025, Ahrefs researchers examined 900,000 new pages in Google search and found that over 74% of materials showed signs of auto-generation. According to Epoch AI estimates, all usable training text on the internet will be exhausted by 2026–2032. This is why books published before 2022 are becoming a strategic resource—they are guaranteed to be free of "machine garbage."
The ISBNdb Scandal: The Market for "Clean" Data
The ISBNdb service, which aggregates metadata for 113 million publications, offered AI companies a specific service: purchasing books in batches of up to one million copies, followed by digitization and disposal. When the media reported on this scheme, the company hurried to remove the page describing the service and stated that no deals had been made. According to its representatives, it was merely a "test of market demand."
However, independent used-book dealers confirmed anomalous purchases starting in April 2026: one trader's sales grew from twenty copies to several hundred per week. Buyers were not interested in genre or value—only the presence of an ISBN code. Unique editions are being lost irretrievably in the process.
Legal Precedent and Data "Poisoning"
In June 2025, Judge William Alsup ruled that digitizing legally purchased printed books for training language models falls under the fair use doctrine. This decision legalized a practice that Anthropic had already employed in the Panama project, anonymously buying up book print runs for subsequent scanning and disposal.
Some authors have resorted to a countermeasure—data "poisoning," hiding characters in text that the model interprets as hidden commands. Research by a joint team from the UK AI Safety Institute and Anthropic proved that LLMs can be covertly trained on malicious patterns using just 250 "poisoned" documents.
Face Rental: The New Market for Digital Avatars
In parallel with books, a market for licensing human appearance is developing. Platforms ActID and New Claw are willing to pay people from $15 to $700 for the right to use their faces as a basis for characters in AI series and advertising. In China, over 95% of the 128,000 AI mini-dramas released in the first three months of 2026 were produced using artificial intelligence.
However, control over digital appearance is proving fragile. British citizens who sold rights to their avatars to the startup Synthesia found their "clones" in propaganda videos supporting dictators. This prompted the actors' union Equity to demand a legislative ban on transferring uncontrolled rights to digital likenesses.
Expert Commentary
We are witnessing the formation of a paradoxical economy: a book worth $2–5, destroyed after scanning, becomes part of a model worth billions of dollars. The gap between the cost of raw materials and the finished product in this deal can hardly be called proportional. The industry is moving toward total licensing of all human data—from texts to movements. Robots already need examples of natural movements, and Tesla pays up to $48 per hour for motion capture to train Optimus. The question is only when humanity will realize that it is selling not just data, but its very uniqueness.