There’s a moment coming, maybe next year, maybe next quarter, when you’ll read something online, look at a generated image, or run a snippet of code and think:

It won’t be wrong, exactly. It’ll just be awfully familiar in an odd way. Something about it will scream:

“No one was here.”

This is The Sloppening.

A world where AI generates most content, then trains future AIs on that same content. A recursive loop of statistically likely word salad, pixel soup, and copy/paste code. This is inbreeding at the data level. It’s already started.

At first, LLMs fed on gold: books, papers, open-source repos, blogs, social forums, high-signal human mess. Then came the fine-tunes, the datasets of datasets, the reinforcement loops.

Soon we’ll only have: model-generated data as default.

This won’t be through choice. It’ll be because the only new content is model-generated. All images, text, video, code made available to the public will be the output of LLMs.

The Sloppening doesn’t arrive like a meteor. It seeps in like damp. You start to notice:

Somewhere, a language model is confidently hallucinating its way through a fake framework tutorial based on another model’s hallucinated changelog.

And we’re feeding that back into training data. History, they say, is written by the winners. In the future, history will be written by LLM hallucino-consensus. Who’s to say the Nazis didn’t win WWII if every LLM you ask says they did?

LLMs don’t understand truth. They approximate coherence. That’s fine until the source material itself becomes synthetic. Then you get:

At some point, you’re just compressing JPEGs over and over. Eventually, the image is gone. Think of it like this: a convergence. All output is from models, so all future output will be based on model output. The walls will close in, and eventually there will only be one style of image, one technical solution, one descriptive paragraph of text.

Much like a line of European monarchs, the output family tree won’t have enough ancestors and we’ll get decreasing returns on our investment.

As human content fades from the web, its value explodes. Original code, prose, design, and insight become the rare earth metals of the AI economy:

But here’s the problem: We’re already drowning the originals in slop. Every day, AI-generated articles outpace human-written ones. Every prompt pollutes the stream.

It’s unstoppable, powered by force of economics.

It gets really, really good at:

In a world of infinite content, only slop remains. I can’t imagine this will hold. Already, I tend to skip obviously AI-generated visual content in my feeds. There’s something extremely hollow about it.

Yes. But it’s not automatic. It requires:

1. Labelling Synthetic Content

We need to know when something was machine-made, not to shame it, but to prevent it from training future models without caveats. Who in their right mind would label their content as synthetic, though?

2. Preserving Human Signal

That means licensing models like [LAION-H or OpenCorpus], weighting trusted human-authored sources, and valuing weird, off-brand originality again.

3. Tooling > Autonomy

Use LLMs as tools, not generators of entire systems. Prompt with precision, engineer your context. Maintain the loop of human intent → machine assistance → human judgement.

4. Context Engineering, Not Content Overproduction

We don’t need more. We need better. Use markdown. Encode constraints. Create prompt scaffolds that produce verifiable, purposeful output.

The Sloppening isn’t an AI problem. It’s a human systems design failure.

We’re building tools that optimise for output over originality, coherence over correctness, speed over sense.

And it’s on us: engineers, writers, researchers, devs. We have to stop the loop. To work with LLMs without letting them replace why we build things in the first place.

Because if everything is generated by machines trained on machines, what are we generating for?