Model collapse is the progressive degradation that occurs when AI models train on data generated by previous AI models. Research published in Nature in 2024 established the phenomenon; follow-up work in 2026 confirms it is already happening and offers partial mitigations. Meanwhile, “AI slop,” low-quality mass-produced AI content, was Merriam-Webster’s 2025 Word of the Year and now accounts for a substantial share of new web content. The combination explains why frontier labs are going to unusual lengths, including buying and destroying physical books , to secure human-written training data.

What happened

Shumailov et al. published “AI models collapse when trained on recursively generated data” in Nature in July 2024. The paper demonstrated that when models train on outputs from previous models, the tails of the original distribution disappear: rare, surprising patterns get smoothed out, diversity drops, and performance degrades over generations. The researchers called this model collapse and showed it affects variational autoencoders, Gaussian mixture models, and large language models alike.

In July 2026, researchers at UC San Diego and Stanford published ForTIFAI in npj Artificial Intelligence, introducing Confidence-Aware Loss functions that can delay collapse onset by over 2.3 times. The practical finding: down-weighting high-confidence tokens during training filters out likely machine-generated artifacts.

The problem is already in motion. According to estimates cited in ACM Communications, over 50% of new web content is now AI-generated. Projections suggest most available new training data could be machine-generated by 2030. The recursion has started: the training data for tomorrow’s models is already contaminated by yesterday’s outputs.

AI slop entered the dictionary as Merriam-Webster’s 2025 Word of the Year, defined as low-quality AI content mass-produced with minimal curation. A June 2026 report from Columbia’s Institute for Global Politics documented how AI slop threatens research integrity and information ecosystems. The NIH published similar findings on the “vicious spiral” of uncurated machine-generated content in scientific publishing.

Why it matters for builders

Three practical implications. First, data provenance is now infrastructure. You need to know where training data came from, how it was generated, and whether it has been through a model already. Detection tools are in an arms race with generation quality and are losing. Watermarking has no universal standard and can be stripped. The metadata layer of the internet was never designed for this.

Second, scaling alone does not fix the problem. Throwing more compute at a degraded dataset amplifies the degradation. The marginal gains from scaling are shrinking while the data supply becomes increasingly circular. Benchmarks measure capability on narrow tasks, not distributional richness, so they mask the underlying decay.

Third, the hunt for pre-AI human-written data is real. Anthropic’s Project Panama bought and destroyed millions of physical books because, in internal documents, books represent “curated, peer-reviewed, domain-specific human knowledge” that “no web crawl can replicate.” ISBNdb now markets bulk book acquisition to AI developers specifically because older print predates the AI-generated flood.

If you are fine-tuning or building your own models, audit your data sources. If you are building products on frontier models, understand that the models themselves may be showing early signs of homogenization, outputs that read fine but carry less information per sentence, the same hedging phrases, the same predictable cadences.

Sources

Further reading