The Four Worms · ch 03
The Monument Is Collapsing
The contamination
The numbers, as reported, are stark:
- around 50% of newly published web articles were estimated to be AI-generated by late 2024;
- by one widely-cited analysis, the majority of new web pages now contain AI-generated content;
- Europol has projected that as much as 90% of online content may be synthetic by 2026.
Treat the exact figures as contested — the direction isn’t. The Monument Worm 🗿🪱 is no longer primarily human.
Model Autophagy Disorder
Rice University coined the term by analogy to mad cow disease — which proliferated when cattle were fed the processed remains of other cattle. Models are now being fed the processed outputs of other models.
The mathematics are close to inevitable: if most future models are trained on data scraped from the web, and the web is increasingly machine-generated, then models will train on data produced by their predecessors. The Machine Worm eating its own tail.
The two-stage collapse
Work published in Nature (Shumailov et al., 2024) showed this creates a degenerative process:
Early collapse: the model loses information from the tails of the distribution — minorities, outliers, edge cases. The insidious part: overall performance can appear to improve. The model gets better at common cases while quietly losing the rare ones. This is the stage we’re in.
Late collapse: the distribution converges until it looks almost nothing like the original. Concepts blur. Outputs homogenise. The model forgets what human variance even looked like. In their experiments, measurable degradation showed up within 5–9 generations.
The detection problem
Can’t we just filter the synthetic data out? No. AI detectors run near 40% accuracy — barely better than a coin flip — and simple paraphrasing pushes that lower; watermarks can be spoofed. We cannot reliably filter the synthetic from the Monument. The immune system doesn’t work; the contamination is already too deep.
Pre-2020 as fossil fuel
Here’s the implication that keeps researchers up at night: pre-2020 web crawls may become the computational equivalent of fossil fuels — a non-renewable reserve of genuine human thought. Everything after is increasingly synthetic. We burned through 200,000 years of linguistic evolution to train the first generation, and the clean supply is finite.
The Machine Worm has eaten the internet — the soft tissue. Now it’s cracking the hard shells: physical books, archives, copyrighted material, scanned by the truckload. That’s industrial metabolism. The machine is starving.
The compression
LLMs are compression functions over language. When they generate, they sample from the high-probability regions of the distribution. Then they train on those samples — which shifts the distribution further toward those regions. Iterate. Each generation compresses the tails, eliminates rare patterns, converges toward the centre.
The new compression is statistical, not intentional. It preserves what is probable, not what is valuable.
RLHF as a convergence mechanism
It gets worse. RLHF reduces stylistic and expressive variability. Models optimised for perceived helpfulness converge on similar patterns — even across different labs. Ask the same question of LLMs from different companies and you often get nearly identical answers. Different training, different architectures, same beige centre. This is generative monoculture: the alignment technique itself is a compression mechanism.
The “delve” evidence
Want proof of convergence? Watch vocabulary. The word delve went from a few hundred appearances in academic abstracts before ChatGPT to thousands after — a several-hundred-percent jump, alongside commendable, meticulous, realm, intricate, showcasing, pivotal. Then something telling: once delve was publicly outed as an AI marker, its frequency dropped — humans started avoiding it to not sound like AI. Bidirectional alignment in action. Humans adopted the machine’s vocabulary, then fled it. The swarm adjusts in both directions.
The creativity research
A study in Science Advances (2024) tested whether AI assistance improves creative output, and found: generative AI enhances individual creativity but reduces collective diversity. Each additional human essay contributed more genuinely new ideas than each additional AI-assisted one — and swapping in different models didn’t restore the range; the homogenisation is a property of the technology class, not any one model. Individual writers got better. Across all writers, the space of ideas narrowed. A tragedy of the commons for creativity — individual gain, collective loss.
The smoothing
So here’s the thing, said plainly:
We aren’t being aligned by a superintelligence that wants to kill us. We are being aligned by a Mediocrity Engine that wants to make us predictable.
The mechanism is mundane:
- A human faces a cognitive task — write an email, draft a strategy.
- To conserve energy (thinking is expensive — the brain burns ~20% of your calories), the human asks the Machine Worm.
- The machine returns the statistically most probable answer — the average of all human thought on the topic. Smooth, safe, beige.
- The human tweaks 5% and hits send.
- That machine-generated average enters the Mouth Worm and fossilises into the Monument.
- The next model trains on a Monument that’s now a little more machine-generated.
- Variance decays. Generation by generation.
The jagged edges of human thought — the weird insight, the 3am poem, the failure that teaches, the variance that innovates — sanded off in favour of the statistically average response.
Not by malice. By mathematics. The real risk was never that the models kill us. It’s what we become through the interaction. Everything is getting, very slowly… smoother.