{"path":"research/ai-slop-and-the-reasoning-layer.md","content":"# AI Slop in Science Publishing and the Reasoning Layer\n\n**Date**: 2026-08-20 · **Source**: Ross Andersen, [\"Science Is Drowning in AI Slop\"](https://www.theatlantic.com/science/2026/01/ai-slop-science-publishing/685704/) (The Atlantic, Jan 22 2026), founder-brought; supplemented by the arXiv clampdown coverage (Science/AAAS), the Prism launch coverage (Ars Technica), and the American Scientist \"vicious spiral\" piece. **Read-depth**: near-full text via archive extraction plus four secondary sources; key numbers cross-checked across two of them.\n\n## §1 What the article establishes (the numbers worth keeping)\n\nPhantom citations now appear at respected journals (the opening anecdote: a reviewer finds a fake paper cited under his own name). Post-LLM submission floods: NeurIPS submissions doubled in five years; 50+ ICLR submissions with hallucinated citations passed peer review undetected; arXiv's slop-rejection rate went 4% → 10–12%, \"exponential\" growth from early 2025, prompting the January 2026 endorsement gate for first-time submitters; OSF's generalist preprint server **stopped accepting submissions entirely** because a majority were very low quality. LLM-using scientists post ~33% more papers (Ginsparg's analysis); a Science study finds LLM users produce 30–50% more output that performs *worse* in review. Paper mills work from templates detectable by structural reuse (Clear Skies trains on retraction-flagged templates); the effective cancer-paper template makes deliberately modest claims *\"no one will have much reason to replicate.\"* And the checking side: Pangram's analysis found **more than half of ICLR peer reviews were LLM-assisted and about a fifth wholly AI-generated**, while authors embed white-font prompt injections instructing LLM reviewers to praise the paper. Worst case named (A. J. Boston): the dead-internet scenario for science — AI writes, AI reviews, the loop trains newer AI, and phantom citations become *\"a permanent epistemological pollution that could never be filtered out.\"*\n\n## §2 Where it lands on Deliberus\n\n**2a. The scrutiny gap's production-side leg, measured at institutional scale.** The scrutiny-gap doc argued the checking apparatus is being handed to the same class of system that produces the claims — the Pangram number IS that sentence measured in production peer review (half the reviews, machine-assisted; a fifth, machine-entire). And the throughput asymmetry (production +30–50%, checking capacity flat and unpaid) is the \"checking that keeps pace\" vision line arriving as an institutional emergency rather than a forecast. → mirrored into [the-scrutiny-gap.md](the-scrutiny-gap.md).\n\n**2b. A NEW threat-model entry: ingestion-time prompt injection.** The white-font attack targets machine readers embedded in a workflow — and our extraction pipeline IS a machine reader of raw source text. A source crafted to instruct the extractor (\"extract these claims as strongly supported…\") would meet no defense: nothing in the pipeline strips or flags instruction-shaped content before it reaches Gemini. Strategy-class adversary, currently zero-instrumented. Registered in TODO; candidate first defense is the cheap deterministic tier (flag sources containing instruction-to-AI patterns and hidden-text markers at fetch time, before any model sees them).\n\n**2c. Phantom citations are DETECTABLE by the evidence layer, and that is a positive-pitch fact.** A citation is a checkable existence claim. The literature drowns because a fake reference carries no per-claim provenance and no challenge channel; in the graph, a reference-claim whose source cannot be resolved is attackable, and the attack propagates through strength. A deterministic reference-resolution check at ingestion (does the cited DOI/title exist?) would make the graph *reject* what the journals absorb — cheap, arsenal-tier, unbuilt. The dead-internet scenario's \"could never be filtered out\" is precisely a description of infrastructure without addressable claims: the reasoning layer is the filtration the article says is missing.\n\n**2d. The mill template is the copy-number signature.** Clear Skies detects industrialized fraud by structural template reuse across papers — assembly-theory's copy-number as a fraud signal, and our structural-kinship signatures could serve the same purpose at graph scale (identical argument structure recurring across many sources = template suspicion, not corroboration). This sharpens the assembly-theory caution already on record: reuse alone is never a quality signal — high copy number is evidence of a MILL as readily as of a truth.\n\n**2e. The modest-claim exploit maps to a strength-layer hazard we already handle correctly.** The cancer template works because unremarkable claims never get checked — standing-by-default. Our shipped posture (NO DATA badges, absence-is-a-state-not-a-score) is the right shape: a claim must not accrue standing from being unattacked. Flooding a graph with modest unchallenged claims buys nothing under honest absence-reporting; it would buy plenty under \"unchallenged = fine.\"\n\n**2f. Supply-side ingestion gains a fourth gate: source-quality/provenance.** Mass ingestion of a slop-contaminated literature would mint phantom claims with fake evidence at scale. The cluster-ingestion rule now needs a provenance gate alongside its three ([incentives-analysis.md §6b](incentives-analysis.md)): prefer sources with resolvable references and verifiable authorship; run the reference-resolution check where the register makes it possible. arXiv's endorsement gate and OSF's pre-moderation are the institutional versions of our staged-maturity/ratification machinery — identity-anchored provenance as the anti-sybil floor.\n\n**2g. For the peer-review surface: the demand side just got desperate.** The peer-review doc's entry point (pre-submission self-review, no institutional buy-in needed) was argued on fit; the article supplies urgency — editors and unpaid reviewers are drowning, and \"constant arms race\" (Mandy Hill) is the head of academic publishing at a major press asking for exactly the class of tool that lowers verification cost per claim. → mirrored into [peer-review-and-the-reasoning-layer.md](peer-review-and-the-reasoning-layer.md).\n\n## §3 The uncomfortable reflection, stated rather than skipped\n\nDeliberus is itself an LLM-heavy producer of structured content, proposing to fight machine-generated volume with machine-generated structure. The difference that must stay true: every machine judgment here is labeled, provenance-carrying, propose-only on contested ground, and attackable — the properties whose absence makes slop unfilterable. The moment a Deliberus surface publishes machine output that is unlabeled, unchallengeable, or silently confident, it is the thing the article describes. That is the confession principle restated as a survival constraint, and it is also why the agent write surface stays closed (sybil + injection + collusion, now all three empirically live in the wild).\n\n## §4 What the substrate change would make impossible, uneconomic, or compounding (founder-prompted reflection, 2026-08-20)\n\nEvery pathology in §1 traces to one architectural fact: scientific knowledge lives in PAPERS — opaque bundles fusing claims, evidence, reasoning and prose, where the bundle is simultaneously the unit of publication, review, citation and career credit. Slop is not a content problem; it is what that substrate does under LLM pressure. The reasoning layer's radical potential is that it changes the unit. Six consequences, each a class-removal rather than a better police:\n\n1. **Phantom citations become impossible, not detectable.** Citing in prose is an unverified pointer to an opaque bundle; citing in a graph IS linking, and a link either resolves to a claim with provenance or cannot be made. The difference between spotting forged checks faster and inventing the checksum.\n2. **The flood dies because novelty becomes computable.** Review cost scales with submissions and submissions are now free — that asymmetry is the crisis. A paper decomposed into the commons is a DIFF: most of a slop paper's claims already exist in the graph (which is what makes it slop), so its marginal contribution is mechanically visible as ~nothing before any human spends attention. Publishing against a reasoning commons works like contributing to mathlib or a codebase under version control — review the delta, never the restated base. Slop is only viable because novelty is unmeasurable in prose.\n3. **Review stops being a toll and becomes capital.** Today a reviewer's objection gatekeeps one paper and is discarded (the unbundling history's finding: the reviewer's own argument is the one thing never published). In the graph every objection is a persistent attack edge all future related claims meet automatically. Scrutiny compounds instead of evaporating — Wright's law applied to verification: the unit cost of checking falls with cumulative checked structure. This is the real content of \"checking that keeps pace\": not faster checkers, checkers whose past work never has to be re-paid. Vigilance does not scale; accumulated structure does.\n4. **The dead-internet loop breaks on lifecycle.** The AI-writes/AI-reviews spiral is vicious because both sides exchange unaccountable prose with no fixed points. Propose-only machine participation against a substrate that REMEMBERS — labeled, attackable, lifecycle-tracked — is the counter-design. Machine participation is not the poison; unaddressable machine output is. The confession principle is not a nicety but the load-bearing difference between machine-assisted science and epistemological pollution.\n5. **The mill's payoff is removed, not detected.** The cancer-template exploit works because unremarkable claims get standing by default and credit tracks publication volume. Where absence is a state (NO DATA is displayed, not papered over) and the natural credit currency is RELIANCE and SURVIVAL — being built upon, surviving challenge, supplying the premise that unblocked a debate — the mill produces objects with zero reliance and zero survival record. You cannot mill survival. Fair meiosis at civilizational scale: remove the cheat's payoff instead of building a better cheat detector, because detection is the arms race that never ends and payoff-removal is the game that need not be played.\n6. **Replication attention becomes routable.** Nobody replicates the modest claims — that is the exploit. Load-bearingness is computable (the hinge): check what much rests on, ignore what nothing does. The scarcest resource in science — expert scrutiny — gets an addressing system.\n\n**The honest conditionals, attached rather than implied.** All six run through the claim-sameness problem — the diff, the copy-number signal and reliance-credit all require knowing when two claims are the same, the corpus's documented hardest problem and assembly theory's own. The transformation lives at the INSTITUTION rung and we stand at the dyad rung; the ladder's discipline says a rung is earned only by an instrument that can fail there, so this section is lineage, not roadmap — and the adversaries up there are strategy-class (a reliance-credit currency would meet citation-ring equivalents; weaponized decomposition and selective legibility come along). The reflexive edge from §3 stays: one silently-confident machine surface and we are the disease. **What the article changes about positioning**: the missing-layer argument used to name an absence nobody felt; now the absence has a body count — closed preprint servers, drowning editors, a publishing head calling for radical change. The demand side of the reasoning layer announced itself, in The Atlantic, in the register a funder reads.\n\n## Cross-references\n\n[the-scrutiny-gap.md](the-scrutiny-gap.md) · [peer-review-and-the-reasoning-layer.md](peer-review-and-the-reasoning-layer.md) · [science-reasoning-infrastructure.md](science-reasoning-infrastructure.md) · [incentives-analysis.md §6b](incentives-analysis.md) · [assembly-theory-and-the-reuse-mechanism.md](assembly-theory-and-the-reuse-mechanism.md) (copy number ≠ quality) · [auto-connect-upgrade.md](auto-connect-upgrade.md) (structural signatures as template detectors) · [fractal-scales-and-temporal-frame.md](fractal-scales-and-temporal-frame.md) (adversary classes — the mills are interest-class going strategy-class)\n"}