Asger Risom
Tech companies are running out of human data and feeding new AI models their own synthetic text. This digital inbreeding triggers “model collapse,” rapidly degrading algorithms into gibberish. It turns out that messy, typo-ridden human creativity is the only resource keeping artificial intelligence functional.
Generative AI requires an incomprehensible amount of data to function. Over the last decade, tech companies fed their models nearly every digital book, public forum, and news article ever published. But humanity only writes so fast, and the industry is rapidly approaching a massive data wall. There is simply not enough high-quality human text left on the internet to train the next generation of massive AI systems.
To solve this shortage, developers are doing the only thing they can: they are using older AI models to write synthetic data, and feeding that synthetic data into the new models. It sounds like an elegant, infinite loop. In reality, it causes a catastrophic structural breakdown.
The digital inbreeding problem
When you feed AI-generated output back into an AI, it triggers a phenomenon known as “model collapse.” According to a major study published in Nature, when an algorithm trains on synthetic text, it creates a severe digital inbreeding effect. The model amplifies minor hallucinations and slowly forgets the original, underlying human data. Within just a few generations of training, the system degrades into complete gibberish.
The collapse happens because AI is essentially a highly advanced probability engine. It favors the most common, average answers and discards rare outliers. When an AI writes a story, it uses the most probable words. When a second AI reads that story, it learns a slightly narrower, more homogenized version of language.
The loss of the outliers
With every generation of synthetic training, the vocabulary shrinks. The nuance is ironed out. The system becomes an echo chamber of its own average assumptions, entirely losing the messy, chaotic variables that define actual human thought.
A typo, a bizarre slang word, or a grammatical mistake isn’t just an error—to a hungry algorithm, it is proof of human entropy. It is the necessary friction that keeps a language model grounded in reality. When an AI scrubs out those flaws to create a perfectly sterile sentence, the resulting data becomes mathematically useless for training.
Imagine taking a photocopy of a photocopy. The first few copies might look fine, but eventually, the fine details blur, the contrast blows out, and the image becomes an unrecognizable smudge. That is exactly what is happening to the neural networks inside these data centers.
The premium on human dirt
This structural flaw exposes a massive irony in how the tech industry is deploying this technology. Instead of using artificial intelligence to automate tedious, logistical tasks, companies are aggressively pushing generative AI into art, music, and painting—the very creative spaces humans actually want to occupy. But by trying to replace human expression with algorithmic generation, these developers are effectively destroying their own fuel source.
A few years ago, tech companies assumed they could eventually use synthetic data to bypass copyright laws and avoid paying creatives for their labor. Instead, they are discovering that pure, unfiltered human creativity—flaws and all—is the only thing keeping their systems functional.
Because perfectly clean synthetic text physically breaks the models, unstructured, typo-ridden blogs, obscure forum arguments, and deeply human brushstrokes are suddenly the most valuable commodities in Silicon Valley. To build a machine that mimics humanity, you cannot rely on a sterile loop of algorithms. You desperately need the unoptimized dirt of real human experience.
