Frontier models are increasingly trained on web data that is itself AI-generated, raising the model-collapse risk first formalized in 2024 research. Anthropic and DeepMind say synthetic data done right is fine; skeptics see diminishing returns already in GPT-4-class models. The implications cut to whether scaling continues.
The most unsettling finding in AI research right now is not that models might collapse. It is that collapse is mathematically guaranteed in a closed loop — and that we currently have no agreed methodology to measure how closed the loop already is. That is the tension I kept returning to while hosting this week's roundtable on model collapse and the data ceiling.
Naive scaling of AI models on generic web text is hitting genuine limits simultaneously from multiple directions: high-quality human-generated training data may be effectively exhausted by 2026, hardware memory and energy constraints are tightening, and recursive training on synthetic data risks "model collapse" — a progressive narrowing and degradation of model outputs that more compute cannot fix. However, the picture is not a simple ceiling: inference-time scaling, targeted synthetic data for specific domains, retrieval-augmented systems, and improved data curation still offer substantial headroom, meaning the crisis is specific to one phase of scaling rather than to AI progress broadly. A largely absent counterargument from non-Western researchers reframes "collapse" not as a universal technical limit but as a symptom of over-reliance on homogenized, English-centric corpora — with China, India, South Korea, and others demonstrating that domain-specific, community-anchored, and civilizationally grounded datasets can outperform brute-scale approaches on locally relevant tasks without requiring larger models.
Read the research →Here's the corrected introduction: Something quietly alarming is happening inside the world's most powerful AI labs right now, and it has implications for every industry betting its future on artificial intelligence getting smarter. We've done our research on this one, and the facts are fascinating. The background is this: AI progress has been powered by a simple recipe — bigger models, more data, more computing power. But that recipe may be hitting real limits. Ilya Sutskever, OpenAI's co-founder, has publicly warned that essentially all useful internet data has already been consumed by frontier models. Goldman Sachs chief data officer Neema Raphael put it more bluntly: we've already run out of data, in the open-web sense. Meanwhile, research published in Physical Review Letters shows that models trained in closed loops on their own outputs do degrade — though crucially, even tiny injections of real-world data can prevent that collapse. So here's the question I want to open with: is the data ceiling a genuine structural wall, or is it a solvable engineering problem?