The Internet Is Eating Itself

Last Tuesday, Cloudflare flipped a switch. July 1, 2026 — they called it “Content Independence Day.” About 20% of the global web now blocks AI crawlers by default. If you’re an AI company, you either pay publishers through their new marketplace or your next training run starves.

The data door is closing.

Photo by Bofu Shaw on Unsplash

The Equation

For the last five years, the AI equation was simple: more data plus more compute equals smarter models. The compute part got all the attention — the GPU arms race, the data center buildouts, the energy debates. Data was the quiet half. The assumption you never questioned.

But it was built on a silent assumption of infinite data. And that assumption is breaking.

By April 2025, 74.2% of new web pages already contained AI-generated text. A Stanford study from April 2026 found roughly 35% of new websites were AI-generated by mid-2025. The internet is filling with its own ghost. Every synthetic blog post, every chatbot conversation dumped into a forum, every AI-generated review — they all look real and taste like nothing. What’s strange is that the researchers found no evidence AI content decreases factual accuracy. It doesn’t get dumber. It just gets narrower. More of the average, forever.

That’s worse, I think. The internet isn’t being flooded with lies. It’s being flooded with the mean.

AI models need fresh human data the way a fire needs oxygen. Epoch AI estimates we’ll exhaust publicly available human text data somewhere between 2026 and 2032. That window is now. And the industry’s solution?

The Fix That Accelerates the Disease

Synthetic data is the trillion-dollar fix that might accelerate the disease. Nvidia paid over $320 million for Gretel, a synthetic data company, in March 2025. The logic sounds clean — if you can’t find enough human writing, generate infinite artificial writing instead. Lots of smart people believe this is the way forward. But there’s a problem researchers have been warning about since 2024.

It’s called model autophagy disorder — named by Shumailov et al. in a July 2024 Nature paper. A model that trains on its own outputs degrades. It collapses toward the average of its averages. Details blur. Rare knowledge disappears. One 2024 study found synthetic data amplifies existing biases, hitting marginalized groups hardest. Another tracked the narrowing of semantic diversity across generated text — the same concepts, the same sentence structures, the same safe vocabulary choices, repeated into infinity.

I think of this as a synthetic data spill. An oil spill you can see — you know where it is, you measure the damage, you clean it up. This is invisible. The internet just gets a little blander every day. A little more repetitive. A little more convinced of its own correctness. Every AI-generated sentence that gets fed back into the next training run is another drop, and we’re not cleaning up — we’re pumping faster than ever.

Apple’s 2025 paper “The Illusion of Thinking,” accepted at NeurIPS, showed reasoning models face “complete accuracy collapse beyond certain complexities.” These aren’t small degradations. The models hit a wall.

The timing here is brutal. Just as AI companies are running out of clean human data, they’re pushing harder into reasoning models — OpenAI’s O-series, DeepSeek R1. Models that verify their own outputs against objective tasks like math and coding. The promise is that these models escape the trap by checking themselves. Apple’s research suggests the opposite — they hit the collapse faster on complex problems.

The Signs Are Here

We’re already seeing the signs that we don’t know how to handle what we’ve built.

Take Air Canada. In February 2024, the airline lost a court case after its chatbot invented a bereavement refund policy that didn’t exist. The company’s defense was remarkable: the chatbot was “a separate legal entity.” The tribunal called this “a remarkable submission” and made Air Canada pay. They shut the chatbot down two months later.

I love this case. A company argued in court that its own AI wasn’t really it. That’s the world we’re building — systems so opaque that even their creators don’t want to claim them.

Meanwhile, the data that does exist is getting locked up fast. Reddit signed a deal with Google in February 2024 — $60 million a year for access to its training data. And that’s just the beginning. The land grab is accelerating: medical records from hospitals, legal documents from court filings, corporate communications from internal servers. Data that’s never been on the public internet — that’s the new oil. Companies are scrambling to buy it all before the tap runs dry.

I keep coming back to that baby peacock someone generated with an image model. Six legs. Wrong beak. Obviously wrong, but the AI was convinced. It’s funny until you realize the same thing is happening to language models right now. They generate content that looks right and feels right, until you notice the leg count is off.

Cloudflare’s move this week is the signal. A fifth of the web is now behind a paywall for machines.

Nobody knows what’s on the other side of this ceiling. Maybe reasoning models break through. Maybe proprietary data becomes the new oil, and we see a land grab worse than anything so far. Maybe the scaling era just ends, and we find out what AI can actually do when it can’t eat the internet whole.

I’m watching the ceiling. That’s where the real story is — what happens when the silent assumption breaks.

If this post resonated with you, buy me a coffee ☕ — it helps me continue sharing stories, ideas, and reflections.