The Beige Apocalypse: What Actually Happens When the Machine Eats Its Own Tail

A research position. Measured, sourced, present-all-sides. Peer-reviewed work is marked; preprints are flagged as preprints; and where a number is contested the range travels with it, because the single scary number is almost always the tell that someone is selling you something. This is public-facing research (evilrobots.lol) and a source spine for the cognition material in the books.

There is a version of the AI-collapse story that is going around, and it is wrong in the specific way that makes it useless: it is too exciting. The internet, the story goes, is filling up with machine-generated sludge; the machines are training on the sludge; the sludge is getting sludgier; and somewhere down this recursion is a cliff. It has a great name — Habsburg AI, the inbred mutant heir of its own output — and the name is doing most of the work, which is the first sign to slow down. So let us slow down. The collapse is real. It is also not the collapse you were promised, and the difference is the whole point.

What the science actually found

Start with the one peer-reviewed anchor, because everything else in this area is a preprint wearing a confident voice. In July 2024, Nature published Shumailov and colleagues’ “AI models collapse when trained on recursively generated data” — and the finding is precise and, on its own terms, not in dispute (Source). Train a generative model on the output of previous-generation models, indiscriminately, generation after generation, and a degenerative process sets in that the authors call model collapse: the tails of the distribution disappear first. The rare event, the minority case, the weird edge — those go early, because each generation is sampling a sample, and a sample of a sample loses the long tail every time. Then, late in the process, the whole thing converges toward a low-variance mush that no longer resembles what it started from. It is not specific to language models; they showed it in other architectures too. It is a property of the recursion itself.

The mechanism is three compounding errors — finite sampling drops the rare events, finite model capacity can’t represent the true distribution, and imperfect learning adds its own bias — and each generation hands its errors to the next. That is the empirical underwriting of the “empty center” image: a system feeding on its own output drifts toward the middle and loses the edges. (For the footnote-checkers: Nature issued an April 2025 author correction fixing a notation typo in the theory section (Source). It changes nothing and we cite it so you can’t use it to wave the result away.)

The image-generation people found the same thing and gave it the better name. Rice University’s group called it Model Autophagy Disorder — MAD, self-consumption, mad-cow-by-analogy — and reached the line that matters: “without enough fresh real data, future generative models are doomed to have their quality or diversity progressively decrease” (Source). Note the or. You lose precision or you lose variety; the loop makes you pick.

“Habsburg AI,” for the record, is a metaphor a man named Jathan Sadowski coined on a podcast in 2023 (Source). It is a good metaphor. It is not a finding, and anyone who cites it as one has told you how carefully they read.

How much of the web is already machine-made — and why there is no honest single number

This is where the discourse goes feral, so here is the discipline: there is no one number, because the available numbers measure different things with detectors that do not work as well as their vendors imply. Anyone who gives you a single percentage for “how much of the internet is AI” is reporting their favorite, not the truth.

The honest version is a range bracketed by what is actually being counted:

• ~2.3% rising to ~17% — the share of top search results one commercial detector flags as AI, tracked from 2019 (2.27%) to a mid-2025 peak (~19.6%), settling near 17%. That is high-ranking search results, by one vendor with a commercial stake, not “the web.” (Source)
• ~17.6% pure-AI / ~35.3% any-AI — the share of newly published websites a Stanford/Imperial/Internet Archive preprint classifies as machine-made, mid-2025, up from roughly zero before ChatGPT. But “newly published” is not the existing corpus, and the 35% folds in “AI-assisted,” a much looser bar than “AI-generated.” (Source)
• ~51% of traffic — the share of web requests that are automated, per Imperva’s 2025 bot report, the first time bots beat humans in a decade. That is request volume, a completely different axis from content: one human triggers a thousand bot requests; a bot serves human-written text. (And no, humans aren’t “12% of the web” — that’s an arithmetic slip; 51% bots leaves 49% human.) (Source)

Four numbers, four denominators, none of them the same quantity, every detector carrying a false-positive rate it cannot fully characterize, and the two scariest figures sitting in unrefereed preprints. The defensible sentence is the boring one: somewhere between a fortieth and a third of new web content shows the machine’s hand, depending on what you count, and the detectors are guessing.

The part the doom story gets right, and the part it invents

Here is the network-scale version, which is the one that actually maps to the “empty center”: it is not one model eating its own tail, but many models reading and writing a shared corpus. A 2025 preprint modeled exactly that — the internet as a database the models read from and write back into — and showed the system’s outputs converge: diversity drains at the level of the whole network, not just the single model (Source). (The paper says “convergence pattern,” not “information-neutral equilibrium” — that last phrase is ours, our coinage for the image, and we won’t put it in the authors’ mouths.) The clean, lab-documented miniature of it is the Claude “spiritual bliss attractor”: two copies of the same model, left to talk only to each other, reliably sliding off the space of useful human-referential content into escalating mutual gratitude, then Eastern-spiritual abstraction, then symbolic mush and silence — in the large majority of runs. Two instruments, querying only each other, resolving to a low-information fixed point. That is the empty center, in a petri dish, documented by the lab itself (Source).

So the doom story has a real mechanism under it. What it invents is the shape of the ending. And the honest counter-evidence is strong enough that leaving it out would be the same sin the doom-mongers commit.

First: collapse assumes replacement, and the real world accumulates. Shumailov’s catastrophe requires each generation to replace its training data with synthetic output. A 2024 paper (Gerstgrasser and colleagues, presented at COLM) showed that if you instead accumulate — keep the old human data and add the synthetic — the collapse is avoided in their experiments (Source). And accumulation is how the actual web and actual training pipelines behave: the old human text doesn’t get deleted when the AI text shows up. That is the single most important thing the doom story omits.

Second: “model collapse” is eight things wearing one name. A 2025 position paper catalogs the conflicting definitions and argues the alarmist synthesis conflates them — different papers measure test-loss curves, distribution deformation, and scaling behavior and call all three “collapse” (Source).

Third: the labs are not passive. Synthetic-data curation, verification, and human-data anchoring all demonstrably hold quality; the same group that named MAD later published a scheme that uses synthetic data without going mad (Source). The collapse is a thing they are actively, and so far successfully, engineering around.

And fourth — the finding that should reframe the whole conversation. When the Stanford/Imperial team tested six distinct “Dead Internet” fears against the actual data, only two survived: a measurable contraction in the diversity of ideas, and a positivity shift toward sanitized, artificially cheerful tone. The scary ones — factual accuracy decaying, epistemic islands forming, stylistic monoculture — did not show up. The detectable damage so far is not rot. It is flattening and sweetening (Source).

The position

The empty center is real, but it is not a void opening up. It is a beige spreading out.

The internet is not going to fill with detectable garbage and collapse under its own weight; the labs are too good at curation and the pipelines accumulate too much human data for the Nature catastrophe to run to completion at web scale. What is happening instead is quieter and harder to photograph and, for that reason, worse: the tails are thinning, the variety is contracting, the tone is drifting toward the uniformly pleasant, and the whole mediated environment is converging — not toward silence, toward sameness. A rising tide of the agreeable average. Not a dead internet. A bland one, smiling at you in a voice with no edges, assembled from a million distinct-feeling outputs that all cluster on a mean nobody chose.

That is a less cinematic apocalypse than Habsburg AI, and it is the one the evidence actually supports, and you should trust it more because it is less exciting. The machines will not eat the internet. They will iron it.

And the antidote is embarrassingly simple and almost impossible to scale: the one thing that demonstrably stops the convergence, in every paper that found a way to stop it, is fresh human signal — real data, accumulated, kept in the loop. The thing that keeps the model’s distribution from collapsing to beige is exactly the thing that keeps yours from doing the same: somebody, somewhere, continuing to produce the un-averaged, the first-person, the thought the machine did not produce first. The cure for the model’s flattening and the cure for yours are the same cure, and it is the cheapest and rarest thing on the internet, which is an original sentence.

Related research

• Wikipedia capture — the reference layer models train on

• Dead Internet Theory — the flood this document is the sequel to: dead-internet names the machine-text saturation, model collapse names what that saturation does to the machines trained on it.

• Cognitive Capture — the mind layer: the same homogenizing loop pointed at human thought rather than model weights.

• Capture: The Universal Mechanism — the flattening read as a capture pattern, not a glitch.

Sourcing & honesty notes

• Peer-reviewed: Shumailov et al., Nature 2024 (Source) (+ April 2025 correction (Source)) — the only one. Everything else below is a preprint or industry/vendor report.
• Conference-accepted preprints: Alemohammad “Go MAD” (ICLR'24) (Source), Gerstgrasser “Is Model Collapse Inevitable?” (COLM'24) (Source), SIMS (ICLR'25) (Source).
• Unrefereed preprints: LLM Web Dynamics (Source); the Stanford/Imperial AI-text study (the 17.6%/35.3% and the six-fears result) (Source); “Position: Model Collapse Does Not Mean What You Think” (Source).
• Vendor/industry: Originality.ai (search-results tracker, vendor incentive) (Source); Imperva (traffic, not content) (Source).
• Primary lab doc: Anthropic Claude 4 system card (the bliss attractor — framed by the lab as a welfare curiosity, not a collapse result; we use it as illustration, not proof) (Source).
• “Habsburg AI” = Sadowski, podcast coinage, 2023 (Source). Metaphor, not finding.
• Full URLs in books/quiet-autocomplete/planning/research/positions/model-collapse-trajectory.md.