AI companies are increasingly seeking human data for training models amid concerns about model collapse, prompting legal actions and new licensing markets.
ISBNdb, a service managing metadata for 113 million publications, recently faced backlash for offering bulk book purchases to AI firms for digitization and destruction. After media scrutiny, the company quickly erased the service details from its website, claiming no agreements were in place and that the initiative was merely a "market demand check."
Meanwhile, platforms like ActID and New Claw are compensating individuals between $15 and $700 for the rights to use their images as characters in AI-generated series and advertisements.
This rising demand for "organic" data stems from concerns over "model collapse," a phenomenon where AI systems degrade as they learn from data generated by other models, rather than from human input. This leads to increasingly predictable and average outcomes, detaching from the original data reality that large language models (LLMs) were designed to emulate.
Understanding Model Collapse
The issue of model collapse was highlighted in a study by Ilya Shumailov and his team, published in May 2023, demonstrating that reliance on synthetic data results in models struggling to accurately reflect rare events and subtle patterns, ultimately leading to reduced diversity in responses and an accumulation of errors.
Statistics reveal the gravity of the situation. A study by Ahrefs in April 2025 found that over 74% of newly indexed content on Google exhibited signs of auto-generation. Additionally, independent research by Epoch AI predicts that the pool of text suitable for training could be depleted by 2026 to 2032, making pre-2022 books increasingly valuable as they are less likely to contain synthetic noise.
Some authors are taking countermeasures through "data poisoning," embedding elements in their texts that AI models fail to interpret correctly. Research indicates that even a small amount of this 'poisoned' data can train models on harmful patterns, highlighting the inadequacy of traditional alignment methods in mitigating such vulnerabilities.
Legal Challenges for AI Companies
Some authors are taking legal action against AI companies. In August 2024, Andrea Bartz, Charles Greibner, and Kirk Wallace Johnson filed a class-action lawsuit against Anthropic for copyright infringement related to the training of its Claude models. The lawsuit alleges that Anthropic, under the direction of former Google Books manager Tom Tervey, purchased physical books, scanned them for digital copies, and subsequently disposed of them without public knowledge.
A significant ruling in June 2025 determined that digitizing legally acquired print books for AI training falls under fair use. This has led to practices where purchasing a physical book and retaining the receipt suffices for legal compliance in training models. ISBNdb has reportedly turned this into a service, facilitating the bulk purchase and destruction of print books for AI training.
On July 21, 2026, ISBNdb’s website featured a promotional section offering AI labs the ability to purchase large quantities of printed books, emphasizing that these books are free from AI noise and legal risks. The management later retracted this offer, claiming it was a test of market interest, not an actual service.
Emerging Market for Likeness Licensing
In parallel, the market for likeness licensing is expanding, particularly in China, where platforms like ActID and New Claw are paying individuals for the rights to their images for AI-generated content. This trend has emerged in response to the rapid growth of the AI content market, with a Beijing court ruling in March 2026 that using someone's likeness in AI-generated deepfakes without consent is illegal.
New Claw’s operations chief, Long Liu, explained that these platforms provide a way for individuals to earn money while continuing their offline careers. However, legal experts warn that while this market is a step forward, it does not eliminate all risks associated with the unauthorized use of likenesses.
Furthermore, AI companies are seeking more than just images and text. The voice data market is also evolving, with SAG-AFTRA recently ratifying protections against synthetic voice usage without consent. This reflects a broader trend where AI companies are now compensating individuals for their data, yet the control over how this data is used remains a concern.
