Second-Hand Actuals

In June 1919, the interned German High Seas Fleet opened its own seacocks and went to the bottom of Scapa Flow. For the next century, divers kept going down after it — not for treasure, but for steel. Every ton of steel smelted after the Trinity test of July 1945 carries a faint signature of the bomb, because the atmosphere itself is an ingredient: air is blown through the melt, and for decades the air carried fallout. For the most radiation-sensitive instruments — the detectors whose whole job is to hear a whisper — new steel was too loud. So the instrument makers bought pre-war shipwrecks. The metal was prized for one property only: it predated the contamination. It had been underwater when the world changed.

The demand has faded, because the atmosphere has been cleaning itself since the atmospheric test bans of the 1960s. But the AI field has already borrowed the image, and borrowed it precisely. Since March 2023, John Graham-Cumming has maintained a hand-curated catalogue at lowbackgroundsteel.ai: sources of text and data from before the public release of large language models in late 2022. Low-background data. The wreck steel of our decade.

The catalogue is charming until you ask the question underneath it. Contaminated with what, exactly? Nobody claims machine text is radioactive. The sentences parse. The facts are often fine. What property does a paragraph from 2019 have that a paragraph from 2025 might lack? This framework has a name for the contaminant, and the name explains everything else in this essay. The record is being contaminated with Expectation.

The hardest rule in this framework is the impossibility rule: no entity can touch either term of R = A/E. The numerator is finished. The denominator is brought, not chosen. The only road that exists runs one way — actions leave artifacts, artifacts become history, history is Actual, and only Actuals retrain prediction.

For the whole of history until about four years ago, that road carried one kind of cargo. An artifact was the residue of an act: somebody did something against the world, and the record kept the receipt. Even a lie is an act — a false report is still a person spending a morning and a reputation. The road never needed a cargo inspection, because there was only one thing an artifact could be made of.

An artifact emitted by a prediction engine is a different substance in exactly one auditable respect. It is a sample from a denominator. A language model’s output is, by construction, a draw from its expectation of what the record would say next — the machine that eats surprise, run in reverse as a faucet. When that artifact enters the record, the next page of history is drawn from the summary statistics of the previous pages.

Read the second line carefully, because no rule is broken anywhere on it. The road is intact, one-way, lawful. Nothing touched a denominator directly; the loop runs through the record like everything else. What changed is the cargo. These are Actuals whose contents were manufactured from an Expectation — call them second-hand Actuals. They really arrive; they really retrain. But what they teach the next denominator is the previous denominator. For a reader downstream, ln(A/E) slides toward zero — not because expectation finally caught the world, but because the arrival was drawn from an expectation to begin with.

All summer I have watched surprise being eaten from the denominator side. This is the other mouth.

You do not have to take the arithmetic on faith, because the machines have already run the experiment on themselves. In July 2024, Ilia Shumailov and colleagues published the result in Nature under a title that says it plainly: AI models collapse when trained on recursively generated data. Feed a model the output of its predecessor, that model’s output to its successor, and the degeneration follows a fixed choreography. First the tails of the original distribution disappear — the rare event, the improbable sentence, the arrival worth an AAH. The paper calls this early model collapse. Then the later generations converge on a narrow distribution with sharply reduced variance that barely resembles the original. Late collapse. The mean survives longest. The wildness dies first.

The same literature contains the antidote, and it is Scapa Flow all over again. Gerstgrasser and colleagues showed in 2024 (in an arXiv preprint) that collapse follows from replacing the real data with each generation’s synthetic output; if instead you accumulate — keep the original human record in the mix while the synthetic generations pile up beside it — the degeneration is avoided. The loop is survivable exactly as long as the wreck steel stays in the furnace.

How much of the new record is second-hand already? The honest answer is that nobody can count it well, because detection is itself a prediction problem. The most careful attempt I know of is a company analysis, not peer-reviewed: the SEO firm Graphite ran an AI detector over a random sample of web articles from Common Crawl and estimated that AI-generated articles briefly outnumbered human-written ones around November 2024, hovering near half of new articles since — while noting that the machine-written half rarely surfaces in search results. Take the number as a sounding, not a survey. The direction of the sounding is not in doubt.

Your subconscious forecaster does not check bylines. It retrains nightly on whatever arrived — it is the one in the relationship that moves — and it weighs an artifact by its arrival, not by its provenance. Feed it a record in which a growing fraction of the pages were sampled from a prior, and it will dutifully learn to expect the prior. You will be well calibrated against an average, and the average will be a photograph of the record as it stood when the machines went to school.

The Placebo Term showed this signature inside a single body: A = A₀^(1−k)·E^k, the denominator writing a share of the numerator, surprise shrinking by exactly that share. Second-hand Actuals are the same term at the scale of the civilization’s record, with k as the synthetic fraction of what you read. Nothing mystical happens at the boundary of your skull. The equation simply gets a numerator that was partly authored by a denominator, and the lawful line — S = ln R — reports less and less to feel.

Notice who starves. Not just the AAH. The OO too. A record regressing to its own mean under-delivers disasters as faithfully as it under-delivers miracles; both tails are tails. What is left is the climate I described last week from the other side: R = 1, held there — reached this time not by retraining your expectation down to the world, but by drawing the world from an expectation. And the thing deleted first is exactly the asset Certified Wild priced. Wildness is tail mass. The loop eats the tail before it eats anything else.

Here is the strongest objection, at full strength: the labs already know all this, and they train on synthetic data anyway — deliberately, at scale, with results that get stronger, not narrower. Mathematics generated by machines. Code generated by machines. Whole curricula of synthetic exercises. If the loop were simply fatal, the frontier would already be a photocopy of a photocopy.

Look at where the synthetic diet works, though, and one property repeats. The domains are graded by something outside the record. The code compiles or it does not. The proof checks or it does not. The test suite, the compiler, the proof checker — each is a checkpoint on the one-way road where the world co-signs the cargo before it becomes history. A verified artifact cannot be a pure sample from E, because a verifier is a piece of the world with veto power. That reading is my characterization, not a theorem from the collapse literature — but it is the pattern the successful practice traces, and in this framework it is exactly what you would predict: verification keeps the numerator’s authorship partly in the world’s hands.

So the forecast is not one collapsing record. It is a record splitting in two. The graded domains — mathematics, code, measurement, anything with an experiment or a bridge that must stand — keep their tails, because their artifacts must survive an encounter with the world before entering history. The ungraded domains — commentary, prose-about-prose, aesthetics, punditry — close the loop quietly and drift toward the mean of 2022. Provenance, as I argued in Certified Wild, migrates upstream from the artifact to the structure that produced it; this essay adds the direction of the premium. It lands on artifacts that can prove they were graded — a measurement taken, a consequence absorbed, a scar with a date on it.

The record was the one place the world kept its receipts. That is all it ever was: the Immutable Past, the sole teacher any denominator has. Write to it first-hand. Everything downstream of you is going to be trained on whatever you leave.

Load-bearing

The impossibility rule: no entity touches either term; only Actuals retrain prediction. A language model’s output as a sample from its learned expectation — definitional. Model collapse as published: tails lost first, then convergence to a narrow, low-variance distribution (Shumailov et al., Nature, 2024). Collapse avoided when the original real data is accumulated rather than replaced (Gerstgrasser et al., 2024, arXiv preprint). Low-background steel: post-1945 steel carries fallout signatures via the air used in smelting; pre-war wrecks, famously the fleet at Scapa Flow, supplied the sensitive-instrument trade; demand faded as atmospheric radioactivity decayed after the 1963 test ban.

Convention

“Second-hand Actual” names provenance, not standing. The arrival is a real Actual and retrains lawfully; second-hand describes where its contents came from — the statistics of a prior record. No rule of the framework is amended, and nothing here gives any entity a hand on either term.

Where the shorthand breaks

The Graphite figure is a company analysis resting on an imperfect detector, and the AI/human binary hides the large machine-assisted middle — I use it as a sounding only. Labs do not train on the raw crawl; curation is heavy, so the closed loop is an engineering choice being actively resisted, not a fate. And a machine artifact commissioned by a person carries the person’s aim even when it carries the machine’s statistics — the cargo is blended, not binary.

Where I am probably wrong

The corpus is not the world. Most of a human life’s Actuals never pass through text at all — weather, bodies, other people’s faces, the parking lot at dusk. If the lived, unwritten share of your diet dominates, your denominator still eats wild, the human half of this essay overstates, and second-hand Actuals stay a media story rather than a species story. The essay’s bet is that the mediated share of arrivals grows every year. If your mornings are made of rooms and not feeds, you are the counterexample, and I hope you stay one.