A hundred billion words a day that’s how much text OpenAI’s systems pump into the world, according to CEO Sam Altman. It’s a figure so vast it’s hard to visualize, yet it hints at a looming technical crisis to which Rockstar Games co‑founder Dan Houser likens to “mad cow disease.” Speaking on Virgin Radio UK, Houser warned that generative AI models are scouring the internet for training material, but the internet is rapidly filling with their own synthetic output. Feed cows to cows, he noted, and you get prions; feed AI to AI, and you risk “model collapse.”
Model collapse, variously called AI cannibalism or model autophagy disorder, is a well-documented phenomenon in machine learning research. Diversity erodes, rare patterns vanish, and outputs converge into repetitive, low-variance sludge when LLMs are trained on their own generated content. Work at Rice University and others demonstrates that image generators develop visual artifacts and shed variation in recursive training loops, while text models shed linguistic richness and amplify biases. This doesn’t happen overnight, but without the injection of high-quality human-origin data, it’s inevitable.
Houser’s skepticism goes beyond the technical. He believes generative AI will “do some tasks brilliantly” but fail at others, a position reinforced by scaling‑law research on synthetic data. Large‑scale experiments involving over 1000 LLMs and 100,000 GPU hours found that pure synthetic datasets-whether rephrased web text or “textbook‑style” generated content-consistently underperformed natural web corpora such as CommonCrawl. Optimal mixtures, such as ~30% high‑quality rephrased synthetic data blended with 70% natural text, accelerated convergence without triggering collapse. But as model sizes grow, tolerance for synthetic ratios drops, underscoring Houser’s point: AI’s strengths are narrow, and scaling it indiscriminately invites self‑poisoning.
This contamination risk is compounded by the sheer scale of synthetic content online. NewsGuard has identified thousands of AI‑generated news sites; and synthetic reviews, social posts, and even academic papers are proliferating. With no foolproof detection methods-watermarking text remains unreliable-this “AI slop” seeps into training datasets unnoticed. As NYU researchers have shown, high contamination forces models to expend more compute to maintain quality, inflating energy costs and undermining scalability. For frontier systems already costing hundreds of millions of dollars to train, this is a non‑trivial threat.
Houser’s critique cuts to the human layer, too: the executives leading the charge for AI adoption. “Some of these people trying to define the future of humanity, creativity, or whatever it is using AI are not the most humane or creative people,” he said. That echoes broader ethical concerns about leadership in AI development. Scholars like Justin Biddle argue that AI systems are“value‑laden because they’re human creations, meaning the biases, priorities, and blind spots of their makers are baked into their outputs. When those makers lack empathy or creative depth, the systems they build risk amplifying narrow, utilitarian visions of human culture.
From an engineering point of view, the way to avoid AI’s “mad cow” moment is through disciplined data stewardship. Hybrid training pipelines have to filter out contaminated samples, maintain diversity through human-curated inputs, and apply architectural safeguards to resist collapse, such as minibatch discrimination or KL-divergence annealing. Synthetic data has its place-especially in instruction-tuning or domain-specific augmentation-but it needs to be generated in controlled conditions with critic systems and human oversight to catch subtle degradation before it compounds.
For digital culture, gaming, and creative industries, Houser’s warning is more than metaphor. If the next generation of AI‑assisted tools draws from an internet saturated with its own recycled tropes, the result won’t be richer worlds or more inventive narratives. The outcome will be homogenized, risk‑averse output-the algorithmic equivalent of a copy of a copy. In that scenario, the very qualities that make human creativity unpredictable, context‑aware, and emotionally resonant could become the rarest data of all.
