The Real Reason AI Company's Are Tapping the Brakes (Hint: It’s Not Safety)

The AI boom in recent years has relied heavily on an abundant supply of public data, scraping the entirety of the internet for text to train Large Language Models. However, this abundant source of data is now running dangerously low. By 2026-2028, most high-quality, human-generated text is expected to be exhausted according to research institutions. The tech industry is now grappling with a more grim reality: the replacement of the open web with AI-generated, toxic content.

Today's internet is increasingly dominated by programmatic SEO farms churning out AI-generated affiliate articles, LinkedIn feeds brimming with synthetic thought leadership, AI-written code snippets flooding GitHub, AI-generated recipes padding ad pages, and bots arguing with each other. Users may find this frustrating, but for AI research labs, it's an existential crisis.

The AI models are no longer learning from the organic human thought that once filled the web. Instead, they are consuming regurgitated outputs of previous models, leading to a phenomenon known as Model Collapse. This mathematical inevitability causes the generative models to disintegrate, losing rare counter-arguments, obscure domain knowledge, eccentric metaphors, and linguistic oddities found in human language.

In response, major tech giants are racing to find solutions. Watermarking is a prime example, albeit a flawed one. While it may deter motivated hackers, it fails to protect against the sheer volume of AI-generated, machine-generated content flooding the internet. Labs are pivoting towards synthetic data, but it must be tied to mathematical or logical ground truth, as synthetic text alone cannot prevent Model Collapse.

The outcome of this scenario could be a great flattening, where next-generation models are forced to train on a web dominated by 40% to 60% machine-generated slop, resulting in AI systems that become unoriginal caricatures of themselves. The era of the free, open web may soon come to an end, replaced by proprietary, verified data archives. The demand for original human thought will also become more expensive, as the public web becomes toxic for AI training.