Synthetic Data and Training Provenance

Established “Synthetic data” now names two opposite things, and the difference between them is the whole subject. One flow is pollution: model-generated text and images published to the open web without labels and ingested by the next crawl, by accident, at unknown concentration. The other is product: synthetic training data manufactured deliberately, filtered by execution loops and reward models, and mixed into corpora on purpose — 40 percent of the tokens behind Microsoft’s Phi-4, over 98 percent of the alignment data behind NVIDIA’s Nemotron-4 340B. The first flow is feared as a slow poison; the second demonstrably works. What the two flows share is a root problem: nobody can currently certify where any given training token came from.

Established The fear has a canonical citation. In July 2024, Shumailov and colleagues published “AI models collapse when trained on recursively generated data” in Nature: generative models trained on the output of their predecessors, generation after generation, lose the tails of the original distribution and eventually degenerate into repetitive nonsense. The result is mathematically grounded, replicated across model families and modalities, and — as this brief will argue in detail — measured in a training regime that no production laboratory actually uses. Whether it describes the future of the web-trained model or an artifact of the experimental design is a genuine, named dispute, and this brief declares it rather than averaging it.

Established Meanwhile the substrate is measurably changing. A 2025 study of 43,000 English-language web articles found machine-generated pieces outnumbering human-written ones from November 2024; an Amazon audit found machine translation dominating the web text of lower-resource languages; and the Data Provenance Initiative measured 5–7 percent of the tokens in standard open training corpora being placed behind robots.txt restrictions in a single year. The open web is simultaneously filling with model output and closing to model builders.

Frontier Provenance — a verifiable account of what a model was trained on — is the question no institution yet owns. The EU now compels a public summary of training content, courts on two continents have begun pricing training data (a $1.5 billion settlement; roughly $3,000 per book), and a cryptographic provenance standard exists for images but not, in any deployed form, for text. This brief owns the joint question the surrounding briefs each graze: can machine learning keep feeding on a web it increasingly writes itself, and can anyone still say what a given model ate?

Established The Nature collapse result is real and specific. Shumailov, Shumaylov, Zhao, Papernot, Anderson and Gal fine-tuned OPT-125m on the wikitext2 dataset, generated a synthetic corpus of matched size from the trained model, trained the next generation on that output, and repeated. The generation-zero model reached a perplexity of 34; in the regime where each generation saw only its predecessor’s output, later generations drifted 20–28 perplexity points worse, and by generation nine the model emitted the now-famous litany of “jackrabbits” variants. The authors distinguish early collapse — low-probability events vanish first — from late collapse, convergence on a narrow distribution with shrunken variance, and derive the mechanism from three compounding error sources: statistical approximation (finite samples miss the tails), functional expressivity, and functional approximation. Critically, when 10 percent of the original human data was retained in each generation, degradation was modest — a detail the headlines dropped.

Established The phenomenon replicates across modalities and was independently discovered. Alemohammad and colleagues at Rice, in work presented at ICLR 2024 under the name Model Autophagy Disorder (MAD), showed image generators in self-consuming loops losing either quality (precision) or diversity (recall) unless each generation received enough fresh real data. Ali Borji’s 2024 note argues the finding is not exotic at all: iterated sampling-and-refitting loses tails as a matter of elementary statistics, which makes collapse in the replace-everything regime close to unavoidable — and close to unsurprising.

Frontier The counter-evidence targets the experimental design, not the mathematics. Gerstgrasser, Schaeffer and twelve co-authors (Stanford and collaborators, 2024) reran recursive training with one change: instead of each generation replacing its training set with model output, synthetic data accumulates alongside the original real data — which is what an ever-growing web crawl actually looks like. Under accumulation, test error stayed bounded regardless of the number of generations, across language models, diffusion models and variational autoencoders. On their account, collapse is an artifact of a data-replacement assumption that matches no production pipeline.

Frontier The pessimists answered with sharper theory. Dohmatob, Feng, Subramonian and Kempe’s “Strong Model Collapse” (2024) proves, in a high-dimensional regression setting with experiments up to GPT-2 scale, that even a small synthetic fraction — on the order of 1 percent of the training set — can degrade scaling, and that naive real/synthetic mixing does not rescue performance unless the synthetic share asymptotically vanishes; larger models can amplify the damage before the interpolation threshold and dampen it after. The dispute is therefore not “does collapse exist” but “which regime is ours”: full replacement (collapses, replicated), pure accumulation with real data retained (bounded error, replicated), or curated mixing at frontier scale (theory disagrees with theory, and the decisive experiment has not been published). A tie to declare: the accumulation paper and the leading skeptical position paper share authors (Schaeffer, Kazdan, Koyejo), so the optimistic case is partly one group prosecuting its argument across multiple venues.

Frontier The 2025 position paper reframes the discourse itself. Schaeffer, Kazdan, Arulandu and Koyejo count eight distinct and partly conflicting definitions of “model collapse” in the literature, weigh published experiments by how closely their conditions match real training practice, and conclude that the most-publicized collapse scenarios are readily avoidable with data retention and curation, while less glamorous risks — benchmark contamination among them — are comparatively neglected. Their strongest empirical card is the natural experiment already running: frontier capabilities kept improving through 2024–2026 on contamination-aware evaluations even as the web filled with model output. That observation is genuine evidence against fast collapse; it says nothing yet about slow tail erosion, which benchmarks aimed at the head of the distribution would not detect.

Established The contamination of the substrate is now measured, with stated error bars. Graphite’s 2025 study ran an AI detector over 43,000 English-language articles from Common Crawl published between January 2020 and May 2025: machine-generated articles outnumbered human-written ones from November 2024, after which the share plateaued at roughly half of new articles — the authors’ own explanation being that AI-written pages perform poorly in search and so stopped paying. The measurement leans on a detector validated only against GPT-4o (4.2 percent false positives, 0.6 percent false negatives) and excludes human-edited AI drafts, so it is a floor with soft edges, not a census. For the multilingual web the news is worse: Thompson and colleagues at Amazon showed that a large fraction of web content in lower-resource languages consists of multi-way machine translations — low-quality English content mass-translated outward — and concluded that this “raises serious concerns” for any model trained on scraped multilingual text.

Established Deliberate synthetic data, by contrast, has a string of engineering wins. Phi-4 (Microsoft, December 2024) is a 14-billion-parameter model trained on roughly 10 trillion tokens of which about 40 percent were synthetic — around 400 billion unweighted tokens across some fifty generated dataset types, quality-controlled by self-revision loops, execution-validated code and curated seed material (vendor figures). It beat its own teacher, GPT-4o, on GPQA (56.1 vs 50.6) and MATH (80.4 vs 74.6), while still trailing it on HumanEval — the gains concentrate exactly where outputs can be checked. NVIDIA’s Nemotron-4 340B reports that over 98 percent of its alignment data was synthetic, and shipped the generation pipeline under a license that explicitly permits using the model’s outputs to train other models (vendor). The working distinction the field has converged on: synthetic data curated, verified and anchored to fresh real seeds is an asset; anonymous model output ingested from the web is a contaminant of unknown dose.

Established Benchmark contamination is the corner of the problem with the cleanest measurements. Scale AI’s GSM1k team commissioned 1,000 new grade-school math problems matched in difficulty to the standard GSM8k benchmark and re-tested the field: several model families dropped by up to 8 percent on the fresh problems, with a positive correlation (Spearman r² = 0.36) between a model’s likelihood of generating verbatim GSM8k examples and the size of its gap — partial memorization, measured. Frontier models showed minimal overfitting, and the emerging defensive practice is visible in Phi-4’s report: aggressive decontamination plus evaluation on competitions published after the training cutoff, such as the November 2024 AMC-10/12 exams.

Established The stock of human text is large, finite, and now shrinking in accessibility. Epoch AI’s estimate (Villalobos et al., 2024) puts the effective stock of quality-adjusted public human text near 300 trillion tokens (90 percent confidence: 100–1,000 trillion) and projects full utilization between 2026 and 2032 depending on overtraining. Frontier training runs already consume 14–36 trillion tokens. Against that, the Data Provenance Initiative’s “Consent in Crisis” audit of 14,000 domains behind C4, RefinedWeb and Dolma found 5–7 percent of all tokens newly restricted by robots.txt in 2023–2024 alone, 20–33 percent of tokens from the highest-quality head domains fully restricted, and about 45 percent of tokens under some terms-of-service constraint — with the restrictions aimed unevenly (25.9 percent of tokens restricted against OpenAI’s crawler versus 9.8 percent against Google’s, whose crawler doubles as the search lifeline).

Established Courts have begun pricing training data, and the answers do not yet agree. In June 2025, Judge Alsup held in Bartz v. Anthropic that training an LLM on lawfully acquired books is “exceedingly transformative” fair use, while building a permanent library from pirated copies is not; the piracy exposure settled for $1.5 billion — roughly $3,000 per work across about half a million listed works, four times the statutory minimum — with final approval granted in July 2026 and Anthropic obliged to destroy its LibGen and Pirate Library Mirror copies. Four months before Alsup, Judge Bibas reached the opposite pole in Thomson Reuters v. Ross: training a (non-generative) legal-research competitor on Westlaw headnotes was not fair use. Kadrey v. Meta (June 2025) went to Meta on summary judgment, but with the court advertising that a better-developed “market dilution” record might win next time. And New York Times v. OpenAI reached summary-judgment argument before Judge Stein in early September 2026, with a ruling expected within months on whether the case goes to trial in 2027 — the U.S. administration filing in support of the defendants. In the UK, the November 2025 Getty v. Stability judgment resolved almost nothing about training: Getty abandoned its primary copyright claims for lack of proof that training occurred in the UK, and the court held that model weights are not themselves an infringing copy — while leaving the legality of UK-based scraping and training expressly undecided.

Established Regulators moved before the courts finished. The European Commission published its mandatory training-data disclosure template on July 24, 2025 under Article 53(1)(d) of the AI Act: general-purpose model providers placing models on the EU market after August 2, 2025 must publish a “sufficiently detailed summary” of training content — including the top 10 percent of scraped domains by size, licensed sources, user data, and the origins of synthetic data and which models generated it — with legacy models given until August 2027 and enforcement verification beginning August 2, 2026. It is the first jurisdiction to make training provenance a legal artifact rather than a courtesy.

Established Provenance infrastructure exists for pixels, not for tokens. The C2PA standard attaches cryptographically signed manifests recording origin, edits and AI involvement; by 2026 OpenAI attaches C2PA metadata plus SynthID watermarks to generated images, Google verifies Content Credentials across Gemini, Search, Chrome and Photos, and Canon, Sony and Nikon ship provenance workflows in professional camera bodies. But manifests are stripped by screenshots, recompression and most platform upload chains, so absence proves nothing; major social platforms apply their own AI labels rather than preserving credentials; and there is no deployed equivalent at all for plain text — the medium training corpora are made of. On the supply side, rights-clean corpora exist and are small: PleIAs’ Common Corpus offers about 2 trillion tokens of public-domain and openly licensed text with per-document license metadata (programme’s own figures) — five to eighteen times smaller than what frontier models consume.

Frontier The live question is the dose-response curve at frontier scale. Between the replicated extremes — full replacement collapses, full retention stays bounded — lies the regime every production lab actually occupies: enormous accumulated crawls, heavy curation, and an unknown, growing, unlabeled synthetic fraction. Strong-collapse theory says 1 percent contamination can bend scaling laws; the accumulation results say bounded error; no published experiment settles it at the scale where it matters, because the only actors with the compute treat their data pipelines as trade secrets.

Frontier Is degradation already happening in the tails while benchmarks improve at the head? Early collapse, on Shumailov’s own account, eats low-probability events first: rare facts, minority dialects, low-resource languages, unusual reasoning paths. Standard evaluations sample the head of the distribution and would miss this by construction. Nobody currently publishes tail-coverage metrics across model generations; the Amazon machine-translation audit suggests the multilingual tail is already the most polluted region of the substrate.

Frontier How far can verification substitute for provenance? Phi-4’s gains concentrate in domains with an oracle — code that executes, math that checks. Whether synthetic data can carry frontier improvement in domains without verifiers (history, medicine, law, style) is unresolved; Epoch’s caution that synthetic data has “only been shown to reliably improve capabilities in relatively narrow domains” still stands as of its 2024 assessment, and reasoning-model distillation since then has widened the checkable zone without abolishing its boundary.

Frontier Can synthetic content be measured at corpus scale at all? Every published estimate of the web’s synthetic share rests on detectors whose error rates are known only against the specific models they were validated on; detection and generation are adversarially coupled, and watermark schemes for text remain fragile to paraphrase. A measurement instrument with calibrated, model-independent error bounds is a missing scientific result, not an engineering detail.

Frontier Which constraint binds first: tokens or rights? Epoch’s 2026–2032 exhaustion window assumes the stock stays usable; Consent in Crisis measures it closing at 5–7 percent of tokens per year while litigation prices what remains. It is now a serious possibility that legal access, not physical supply, becomes the binding constraint on training-corpus growth — and that the constraint binds unevenly, favoring incumbents with archived pre-restriction crawls and platforms with captive user data.

Speculative The second-order loop runs through people. Humans increasingly read, imitate and lightly edit model prose, so even “human-written” text drifts toward model distributions — recursion by cultural transmission rather than direct ingestion. Documenting this drift, and whether it produces the same homogenization as direct self-training, is an open measurement problem with early stylometric evidence and no settled methodology.

Established No trustworthy detector at corpus scale. The best public synthetic-share measurement validated its detector against exactly one generator (GPT-4o) and concedes that AI-assisted writing falls outside its categories. At the scale of a 30-trillion-token crawl, even a 4 percent false-positive rate misclassifies more than a trillion tokens; every downstream number inherits that fog.

Established No provenance standard for text. C2PA manifests bind signatures to media files; nothing equivalent survives contact with plain text, which is copied, quoted, re-encoded and stripped of metadata as a matter of course. Statistical watermarks for LLM output exist (SynthID-Text has been deployed) but degrade under paraphrase and translation, and no major lab watermarks the bulk text its models emit through APIs into the world’s documents.

Established Rights-clean supply is an order of magnitude short. Common Corpus’ 2 trillion fully documented tokens against frontier appetites of 14–36 trillion means a provenance-pure frontier model is not currently constructible; the gap must be filled by licensed private troves, by synthetic generation, or by exactly the gray-provenance web data the clean corpus was built to avoid.

Frontier Verification does not generalize past checkable domains. Execution loops and reward models are the quality control that makes deliberate synthetic data safe; outside code and mathematics the reward model is itself a learned approximation, and filtering by it risks laundering the generator’s biases into the next generation with a certificate attached.

Established Decontamination is an arms race conducted with n-grams. GSM1k demonstrated that standard string-matching decontamination misses paraphrased and translated benchmark leakage; contamination-proof evaluation currently requires commissioning fresh problems — expensive, slow, and consumed on first use.

Established The economics still favor scraping. A crawled token is nearly free; the settled price of one infringed book is roughly $3,000; licensing markets are bilateral, opaque and available mainly to the largest labs. Until provenanced data is cheaper than litigation risk, provenance remains a compliance cost rather than a product.

Established This topic sits on statistics older than the field that rediscovered it. Collapse dynamics are iterated distribution estimation — the mathematics of what finite sampling does to tails, studied since Galton; the open theory questions (mixing thresholds, scaling-law deformation) are being prosecuted with random-matrix and high-dimensional regression tools.

Established Measurement depends on detection and watermarking research. Corpus forensics needs calibrated classifiers, robust text watermarks, and multi-way parallelism tricks of the kind the Amazon team used — each a live subfield with adversarial pressure on it.

Established Provenance depends on archival infrastructure that mostly already exists. Common Crawl’s dated snapshots, the Internet Archive, national libraries and versioned datasets are the substrate any chain-of-custody scheme would anchor to; the missing layer is signatures and institutional custody, not storage.

Frontier The legal dependency is a moving target. Fair-use doctrine (four live US precedents pointing in three directions), the EU’s text-and-data-mining opt-outs, and the AI Act’s disclosure regime jointly determine which data may be used and what must be revealed about it; every technical architecture in this brief inherits its constraints from how those settle. Benchmark science — the craft of building evaluations that contamination cannot reach, covered from another angle in Intelligence Measurement — is the final dependency: without it, nobody can even tell whether recursive training is helping or hurting.

Frontier The decisive experiment is the crawl-vintage comparison: train two otherwise identical frontier-scale models, one on a web snapshot predating late 2022 and one on a current snapshot, under the same curation pipeline and compute budget, and measure capability, calibration and tail coverage on contamination-proof evaluations. Any frontier laboratory could run it today; none has published one. A null result would largely retire the pollution fear at current concentrations; a measurable gap would convert model collapse from a laboratory curve into a production cost and make pre-2023 archives strategically priceless. Everything else in this section refines that one test.

Frontier Second: the dose-response study at scale. Extend the Strong Model Collapse empirics beyond GPT-2 scale — controlled synthetic fractions (0, 1, 5, 20, 50 percent) under accumulation-with-curation rather than replacement, tracking where the theoretical 1-percent sensitivity survives realistic data practice and where Gerstgrasser-style boundedness wins. This is the experiment that would actually adjudicate the standing dispute, and at sub-frontier scale it is fundable by any national compute program.

Established Third: tail-coverage monitoring, which needs no new hardware. Track rare n-grams, low-resource-language quality, dialect representation and long-tail factual recall across Common Crawl vintages and across successive frontier model releases. Early collapse predicts the tails thin first while head benchmarks hold; the measurement is cheap and nobody publishes it.

Established Fourth: standing contamination probes. Institutionalize the GSM1k design — privately commissioned, distribution-matched twins of public benchmarks, re-run against each model generation with the gap and the memorization correlation reported. One group did it once for one benchmark; it should be boring infrastructure.

Frontier Fifth: the watermark-persistence field trial. Release watermarked model output into ordinary publishing channels and measure surviving detectability in the next year’s crawl — the number that decides whether generation-time marking can ever support crawl-time filtering.

Established Sixth: the policy result already scheduled. The EU’s Article 53 disclosure regime — templates mandatory for new models since August 2025, enforcement from August 2026, legacy models by August 2027 — is a natural experiment on whether mandated transparency yields provenance information anyone can audit, or a new genre of unfalsifiable paperwork. Its first full compliance cycle is running now.

Established Provenance at training scale is a chain-of-custody problem measured in petabytes. A pipeline that could actually certify a corpus needs signed, dated crawl snapshots; per-document source, license and hash metadata that survive deduplication, filtering, chunking and tokenization; and audit logs binding the final token stream to the documents it came from. Every element exists in miniature (Common Corpus ships per-document license metadata at 2-trillion-token scale); none exists at the 10-to-30-trillion-token scale of frontier practice.

Established Deliberate synthetic generation is already a mature engineering discipline. The published Phi-4 and Nemotron pipelines specify seed curation, multi-stage prompting, self-revision loops in which a model critiques its own drafts against rubrics, execution validation for code, reward-model filtering for everything else, and instruction reversal — roughly fifty distinct dataset constructions in Phi-4’s case. The quality-control insight is uncomfortable for the provenance agenda: these pipelines work by making synthetic data better than the web average, which increases the incentive to manufacture rather than to document.

Frontier Detection engineering has to run at ingest, not at audit. Filtering model output from a crawl requires detector ensembles with per-generator calibration, watermark checks, and duplication forensics running inline at crawl throughput — and failing gracefully, since a 4-percent false-positive rate applied naively would discard a trillion legitimate tokens and, given detector bias against non-native writing styles documented elsewhere, discard them unevenly.

Frontier The archival requirement is a freezer, not a firehose. Preserving verified pre-2023 corpora, and continuing to capture provably human text (keystroke-attested journalism, institutional archives, licensed fresh writing) is cheap relative to training compute but requires an institution with a century horizon — a role national libraries could hold and currently do not.

Established This brief is the supply-side twin of the capability briefs around it. Artificial General Intelligence owns the question of whether scaling continues; this brief owns what it scales on — Epoch’s 300-trillion-token stock estimate and the 2026–2032 utilization window are the data-side wall that synthetic generation is being built to climb. Artificial Creativity establishes that machines beat the human average and lose to the human right tail; recursive training is the mechanism by which that right tail — the part models cannot regenerate — could quietly leave the training distribution.

Established Governance and measurement briefs carry the institutional halves. AI Governance covers the regulatory machinery of which the EU’s Article 53 disclosure template is one instrument; Intelligence Measurement covers why evaluations mean anything, of which benchmark contamination — measured here at up to 8 points of GSM8k inflation — is the failure mode.

Frontier Two farther neighbors share the recursion structure. Artificial Scientists proposes models generating the scientific record future models learn from — the same loop with peer review as the verifier; Collective Intelligence studies information ecosystems degrading under their own outputs at the human layer. The technical adjacencies — watermarking, content credentials, dataset documentation, retrieval-based attribution — are each nearer to deployment than any of them is to sufficiency.

Established The EU AI Office is, for now, the only provenance regulator in existence. Its template obliges disclosure of scraped-domain lists (top 10 percent by size), licensed sources, user-data use, and — notably — the origins of synthetic training data and which models generated it, with updates every six months. Whether its verification powers, live since August 2026, produce audits or filings is the institutional experiment to watch.

Established Civil society currently outperforms states at measurement. The Data Provenance Initiative — an academic collective — produced the field’s only longitudinal audit of consent signals across training corpora; Epoch AI produces the standard data-stock accounting; the C2PA coalition (Adobe, camera makers, platforms, news agencies) built the only deployed provenance standard. None has enforcement power; all are running ahead of the institutions that do.

Established Courts are functioning as the de facto data-pricing institution. The Anthropic settlement built, in passing, the administrative machinery of a collective licensing system — a works list, per-work pricing at $3,000, claims administration reaching rightsholders for 99.5 percent of listed works — exactly the apparatus a statutory license would need, assembled by class action instead of legislation.

Frontier The missing institution is a corpus registry with audit rights. Nothing today can answer “was this document in that model’s training set” with authority: labs claim, regulators request, courts subpoena. A registry holding signed corpus manifests in escrow — disclosed to auditors, not to competitors — is repeatedly proposed and nowhere built; the EU template is a summary, not a manifest.

Speculative National libraries are the natural custodians nobody has appointed. The institutions that already hold dated, verified, pre-2022 text at scale are legal-deposit libraries and web archives; converting them into certified suppliers of provenance-clean training data would take a mandate and money, and would give the public sector its only leverage point over the training-data market.

Established Consent has collapsed faster than the data commons itself. The measured surge of robots.txt and terms-of-service restrictions is the web’s creators withdrawing permission at scale — and the withdrawal is self-defeating in aggregate: it cannot reach labs that already crawled, it is honored unevenly, and it starves precisely the open corpora (C4, Dolma) that academic and audit communities depend on, while private data-holders keep selling access.

Established The tails that collapse first are people. Early collapse removes low-probability data: minority dialects, small languages, unusual lives. The Amazon audit shows lower-resource languages already receiving the web’s worst synthetic pollution — mass machine translation of low-quality English — so the communities with the least data face both the dirtiest substrate and the earliest erosion. Distributional harm of this shape will not appear on any English-head benchmark.

Established Compensation now has a market price, and it is low. Three thousand dollars per book, once, settles the copying of a life’s work into infrastructure that may displace its market — four times the statutory minimum and orders of magnitude below what any author would have licensed at. Whether that price becomes the anchor for a functioning licensing market or the ceiling that forecloses one is an open distributive question.

Frontier Provenance is also surveillance-shaped. A web where every text carries signed authorship metadata solves training provenance and creates an attribution infrastructure that de-anonymizes writers; dissidents, whistleblowers and the merely private publish under exactly the anonymity a rigorous provenance regime would end. The C2PA design — optional, strippable — is weak partly because the strong version is dangerous.

Frontier The recurring analogy is low-background steel, and it is worth taking literally. Steel smelted before 1945 is prized because everything after carries bomb-test radionuclides; text written before late 2022 is acquiring the same status — the last large corpus guaranteed free of model output. The analogy’s sting is that low-background steel became less necessary as instruments improved; whether synthetic-text contamination follows that curve depends entirely on whether detection wins its arms race, which is currently unresolved in the other direction.

Frontier First-mover advantage in data may prove more durable than in compute. Shumailov’s own closing warning was that access to pre-contamination crawls and to genuine human interaction data becomes increasingly valuable; compute advantages depreciate with Moore’s-law dynamics, but an archived 2021 crawl cannot be re-created at any price. A handful of firms holding both the archives and the user-interaction firehoses is a plausible long-run concentration mechanism that antitrust doctrine has no category for.

Speculative A majority-synthetic culture is survivable; an unverifiable one may not be. Civilization has always run on mostly-derivative text; the load-bearing quantity is not the human share but the existence of trusted channels — science, law, journalism, archives — whose provenance is auditable. If provenance infrastructure succeeds in those channels while the open web goes gray, the likely equilibrium is a two-tier textual world: certified corpora for institutions and training, an unprovenanced commons for everything else.

Handwave The strong doom scenario skips its own mechanism. The claim that recursive contamination alone degrades civilization’s knowledge substrate assumes model output displaces rather than supplements human writing, that curation fails, that verification stalls, and that institutions never respond — four load-bearing assumptions asserted, in most tellings, as one inevitability. Every measured trend so far shows adaptation: plateaued synthetic share, contamination-aware evaluation, provenance law. The honest long-run statement is conditional, not fated.

These horizons track when the questions this brief owns — contamination dose-response, corpus provenance, data rights — get answered or institutionalized, taking the September 2026 state as baseline.

• 10 yr: Frontier The US fair-use split resolves through NYT v. OpenAI and its appellate wake; the EU disclosure regime completes several enforcement cycles and either produces the first real corpus audits or is quietly gamed; the crawl-vintage experiment gets run (publicly or leaked); licensing markets settle per-token prices for fresh text; provenance credentials become default for professional images and remain absent for bulk text.
• 25 yr: Speculative Either a text-provenance layer (watermarks at generation plus registries at ingest) reaches usable coverage, or training regimes complete the pivot away from open-web scraping toward licensed archives, captive user data and verified synthetic generation — in which case the open web ceases to be the substrate of machine intelligence and becomes its exhaust.
• 50 yr: Speculative Pre-2023 corpora hold archival status comparable to legal deposit collections; models are trained overwhelmingly on model-mediated data under verification chains whose reliability is itself the central scientific question; whether the human right tail still enters the loop depends on institutions built (or not) in the first decade.
• 100 / 250+ yr: Handwave Claims about permanently poisoned or permanently pristine knowledge commons at this range assume stable institutions, stable media and stable minds across a period in which none of the three has ever been stable; nothing measured today constrains them.

• Depends on nothing on this map: synthetic-data contamination and training provenance sit upstream, at the substrate layer, and no other brief’s result must land first. The dependency runs the other way — Artificial General Intelligence inherits its data constraint from here, and AI Governance inherits its auditability problem.
• Requires (not on this map) a detector for machine-generated text whose error bounds are calibrated across generators rather than against a single model; collapse experiments actually run at frontier scale under accumulation-and-curation conditions instead of laboratory replacement loops; a provenance standard for plain text that survives copying and re-encoding the way C2PA manifests do not; a settled legal price — by precedent or statute — for training on in-copyright works, which four contradictory US rulings have not yet produced; a licensing market in which paying for fresh, provenanced text beats scraping it on cost and risk; and a standing supply chain that keeps capturing verifiably human writing at corpus scale as the open web’s share of it declines.
• Enables continued scaling past the 2026–2032 human-data utilization window; benchmark results that can be believed, because contamination is measured instead of assumed; regulatory audits of training corpora that check manifests instead of accepting summaries; and a rights-cleared training economy in which creators are counterparties rather than plaintiffs.
• Adjacent to Artificial Creativity (whose human right tail is the resource recursion spends), Intelligence Measurement (whose instruments contamination corrupts), Artificial Scientists (the same recursion with peer review as verifier), and Collective Intelligence (the human-layer version of a system feeding on its own output).

Handwave “The internet is already ruined; models will inevitably degrade.” The claim circulates widely with the Nature paper as citation, and the citation does not support it. The collapse experiments replace real data wholesale generation after generation; production pipelines accumulate and curate, the regime in which Gerstgrasser and colleagues measured bounded error. The natural experiment runs against it too: capabilities improved through 2024–2026 on contamination-aware evaluations while the synthetic share of new articles crossed 50 percent and then plateaued. Degradation at realistic doses remains unproven in either direction — which is a reason to run the experiment, not to assert the outcome.

Frontier “Model collapse has been debunked.” Equally wrong, in mirror image. The mathematics of tail loss under iterated resampling is elementary and undisputed — Borji’s note argues it is close to unavoidable in its regime — and Strong Model Collapse shows theoretically that even 1 percent contamination can deform scaling under assumptions no experiment has yet excluded at frontier scale. What is contested is the mapping from regime to reality, not the phenomenon.

Established “Synthetic data means free infinite training data.” Phi-4’s 40 percent and Nemotron’s 98 percent are routinely cited this way; both pipelines work because of expensive scaffolding — curated real seeds, execution oracles, reward-model filtering, decontamination — and the measured wins concentrate in checkable domains. Epoch’s assessment that synthetic data reliably improves only narrow domains has been eroded at the edges by reasoning-model distillation, not overturned.

Established “Detectors can tell us how much of the web is AI-generated.” Every headline percentage inherits a detector validated against a handful of generators; the best-documented study reports 4.2 percent false positives against pre-ChatGPT text and was checked against GPT-4o alone, excludes AI-assisted writing entirely, and measures English articles only. Treat all such figures as bounded estimates of a moving target produced by an instrument the target can see.

Established “Watermarking and content credentials will solve provenance.” For images, credentials exist and are stripped by ordinary platform plumbing; for text, no deployed scheme survives paraphrase; and absence of a mark can never prove human origin, because marking is voluntary. Provenance infrastructure raises the cost of casual misattribution; it does not create a verifiable web.

Established “Courts have settled that training on copyrighted work is fair use.” One district judge held training on lawfully bought books exceedingly transformative; another held training a legal-search tool on headnotes infringing; a third granted judgment to a lab while advertising the theory that could defeat it; the largest case argued summary judgment in September 2026 and is undecided; and the UK court decided only that model weights are not copies, expressly reserving the training question. There is no settled rule — there is a distribution of rulings with heavy tails.

Speculative “We hit peak data; scaling is over.” Epoch’s own numbers say full utilization of the public-text stock between 2026 and 2032 with a factor-of-ten confidence interval, before counting repetition tolerance, licensed private text, other modalities and synthetic augmentation. Data exhaustion is a real constraint arriving on a genuinely uncertain schedule; the confident short version is not a measurement.

Speculative “Human-written text will always command a premium.” Plausible, asserted constantly, and dependent on the detection arms race: a premium requires verification, and if provably-human infrastructure (attested capture, trusted institutions) does not get built, human text becomes valuable in principle and unpriceable in practice — the worst of both.