> Paper: AI-Augmented Science and the New Institutional Scarcities > Author: Lauri Lovén > Affiliation: Future Computing Group, University of Oulu, Finland > arXiv: 2605.02566 | 2026-05-04 > URL: https://arxiv.org/abs/2605.02566
1. AI Is Replacing Scientific Institutions, Not Just Scientists
The paper's core claim: AI is not replacing peer reviewers—it is replacing the institutional function of "review" itself.
Mainstream AI economics (Agrawal, Gans & Goldfarb, 2022) holds that AI makes *prediction* cheap while *judgment* remains scarce. This framework works for enterprises: AI predicts customer churn, human managers decide whether to act.
But scientific institutions are not enterprises. Their core product is not research—it is legitimate judgment:
- Universities certify academic judgment
- Journals certify the validity of claims
- Funders certify which proposals merit support
- Academic societies certify professional competence
- Peer review: distributing claims to a few qualified readers
- Authorship norms: attributing scarce cognitive contributions to identifiable individuals
- Reproducibility practices: assuming publication is a probabilistic signal of truth, because direct verification is too expensive
- Grant review: allocating scarce funds via equally scarce reviewer capacity
- AI-assisted writing explodes submission volume; AI-assisted review explodes review volume
- Review quality degrades unless processes are redesigned
- The queue becomes a cat-and-mouse loop; the final signal is worth less than the input signal
- Cost of producing plausible-looking papers → approaching zero
- Cost of verifying correctness → unchanged or rising
- Value of publication as a signal → systematically depreciating
- Our psychological inertia that "published = credible" → unchanged
- Title: AI-Augmented Science and the New Institutional Scarcities
- Author: Lauri Lovén
- Affiliation: Future Computing Group, University of Oulu, Finland
- arXiv: 2605.02566 | PDF
- Published: May 4, 2026 | cs.CY | 7 pages
- Keywords: AI-augmented science, scientific institutions, peer review, reproducibility, content provenance, commons governance, credentialing
- Companion paper: Institutions for the post-scarcity of judgment (arXiv:2604.22966, submitted to Communications of the ACM)
When frontier AI can review manuscripts, draft peer reviews, score grant proposals, and verify numerical claims at near-zero marginal cost, it is competing with scientific institutions for the same functional role.
> "Institutions whose primary output is certified judgment have no fall-back layer when that output becomes abundant."
This is not about how strong AI is. It is about how fragile scientific institutions are.
2. "Judgment Abundance" Inverts 150 Years of Scientific Economics
Modern scientific institutions rest on one premise: qualified human attention is scarce.
> "Each institution encodes a scarcity assumption. When the assumption holds, the institution produces legitimate signal and scales gracefully. When it fails, the institution degrades from the inside, often before anyone notices."
Surface activity continues while the underlying function is hollowed out: submissions grow, reviews happen, citations accumulate, h-indices rise—but certification no longer tracks truth. The reproducibility crisis showed institutions were already hollowing themselves out; AI merely accelerates the process.
3. Peer Review Is Becoming a Cat-and-Mouse Game
Peer review: a non-converging volume problem
NeurIPS, ICML, and ICLR are not anomalies—they are leading indicators.
Authorship: disclosure does not solve attribution
Flagging "ChatGPT was used" is necessary but does not answer "who contributed what." Single-author abstraction has been leaking in AI/ML papers for years.
Reproducibility: an economic inversion
The cost of producing plausible-looking but unreproducible work is falling faster than the cost of verifying it. Publication is ceasing to be a reliable probabilistic signal of truth. AI/ML benchmarks have already absorbed this shock via test-set contamination, prompt-engineered leaderboard climbing, and unreproducible SOTA claims.
Grant review: rhetorical surface decoupled from intellectual depth
AI-assisted proposals converge on the rhetorical surface of "fundability"—right vocabulary, structure, citations—without matching depth. NSF AI Institutes, EU Horizon Europe, and UK ARIA will be exposed first.
> "Each is an institutional failure mode that compounds silently, not a productivity complaint. Treating any as a tooling problem is a category error: the binding constraint is institutional."
4. Four New Scarcities
When AI-generated claims are abundant, four things become scarce and structurally necessary:
Scarcity 1: Verified Signal
A paper whose computational claims have been reproduced carries information a merely "reviewed" paper does not. Verification infrastructure—reproducibility pipelines, provenance chains, certification, adversarial replay—becomes genuine editorial added value. NeurIPS's ML Reproducibility Challenge and MLSys's artifact evaluation are partial versions, but "they were built when the marginal cost of a fabricated baseline was higher than it now is." Open design question: who pays for reproduction, and who certifies it?
Scarcity 2: Legitimacy
Legitimacy behaves like a commons: "a shared stock that depletes when drawn on without replenishment." A journal's brand dilutes with every visibly unreproducible accepted paper. Legitimacy maintenance is the scarce labor—not certification itself. Institutions overdrawing their legitimacy budget do not announce it; they become irrelevant while surface activity continues.
Scarcity 3: Authentic Provenance
Which model produced a claim? What training data? Which prompt chain? Which human decisions shaped the output? Current publication metadata carries none of this. Components exist (model cards, datasheets, W3C-PROV ontology); what is missing is a unified standard for scientific artifacts. C2PA demonstrates the load-bearing-record approach for media; extending it to multi-component scientific outputs is a system research problem the AI/ML community could lead.
Scarcity 4: Integration Capacity—the scarcity AI cannot buy
The least obvious but, per the paper, tightest constraint: how much delegated AI judgment a scientific community can absorb before readers stop trusting its journals, conferences, and grant review. Authors, reviewers, readers, and downstream users (clinicians, policymakers, journalists, industry) each have tolerance curves; institutions are bounded by the minimum.
> "The rate-limiting factor is not computational; it is the community's tolerance for delegated cognition."
> "This is the scarcity no amount of better tooling can buy."
5. Four Reconstruction Moves
Each move targets one new scarcity, all within the AI/ML community's technical capacity:
1. Reproducibility-first editorial action: a subset of accepted papers carries a "we reproduced the computational claims" statement. What is missing is editorial weight, not technical capability. 2. Tiered certification: separate claim-level checks (numerical consistency, citation integrity, code runs as described—AI-assisted), result-level certification (is the claim interesting—human), and framework-level certification (is the framework sound—human). A consortium of major AI/ML program committees is the natural coordinator. 3. Provenance attestation by default: every AI-augmented paper carries machine-readable provenance—model versions, data sources, prompt traces, human decision points, audit logs. What is missing is institutional commitment plus a hosting layer (arXiv, OpenReview, or community-governed commons). 4. Verification as a commons: verification as a product inherits foundation-model market centralization; verification as a commons does not. Multi-centered governance—many overlapping certifiers—is the most capture-resistant architecture. Learned societies (ACM, IEEE-CS) or funder coalitions are natural conveners; science-led verification commons would complement, not compete with, EU AI Act Article 50 compliance.
Fairness warning:
> "These costs fall asymmetrically. Reproducibility infrastructure, provenance engineering, and tiered editorial labour are paid disproportionately by under-resourced institutions, by the Global South, and by disciplines without AI/ML's tooling base."
6. A Feynman-Style Judgment
Feynman said: "Science is the belief in the ignorance of experts." In this framework, the gap between "knowing a paper was published" and "knowing a paper is correct" is being systematically amplified by AI:
The paper's sharpest line:
> "Producing competent-looking judgment is now the cheapest part of science; producing legitimate judgment is becoming the most expensive."
> "The journals and conferences that survive the next decade will not be those that automated their review queues fastest, but those that rebuilt their verification commons first."
7. Takeaways
Ask yourself:
1. Is my field's certification infrastructure silently degrading—surface activity continuing while the underlying function hollows out? 2. Where is my community's tolerance ceiling for delegated cognition? Have we crossed it? 3. Of the papers I read, how many were "reviewed" versus "reproduced"? Is that gap widening? 4. If AI can produce plausible-looking judgment at zero cost, what is a reliable signal of what is actually correct?
AI's greatest threat to science is not replacing scientists—it is removing the reason scientific institutions exist. When everyone can produce outputs that look like science, genuine scientific output must be certified through new institutional mechanisms. In an era of judgment abundance, legitimacy is the scarcest resource.
---