The recipe for Neapolitan-style pizza dough seemed simple enough. But the first Google result led to a blog post with a photo of a perfect crust, blistered and leopard-spotted. The instructions were odd. They called for an impossible amount of yeast and a baking time that would turn a pizza into charcoal. The second link was nearly identical, as was the third. These were not human errors. This was the internet’s new wallpaper: a layer of synthetic, plausible-sounding nonsense generated to capture clicks.
For years, we treated the internet as an infinite library of human experience, a messy but bottomless well of data to train our artificial intelligence models. That era is over. The well is being poisoned. The very tools built to learn from that library are now flooding it with their own garbage, creating a self-referential loop that threatens to stall the entire AI project.
This isn't a theoretical problem. Researchers call it "model collapse." When an AI model is trained on data generated by another AI, it begins to amplify the errors and biases of its predecessor. Details get smoothed over. Outliers and creative leaps—the very essence of human ingenuity—are discarded in favor of the statistical average. The model learns from a copy of a copy of a copy, and each generation gets fainter. Eventually, it forgets reality and only remembers the echo of its own voice.
What was once a slow trickle of AI-generated content has become a firehose, thanks to the democratization of powerful open-source models. A year ago, producing convincing text or images required access to the expensive, closed systems of a few big labs. Today, anyone with a decent graphics card can download a model like Llama 3 or Stable Diffusion and churn out a million blog posts or a billion images for the cost of their electric bill. Content farms that once paid humans pennies per word now pay nothing.
The result is a digital landscape where the synthetic is indistinguishable from the authentic, and the sheer volume