Epistemic Infrastructures of Science in AI Era Should Rebalance Costs of Generation and Verification
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: We're looking at the paper: "Epistemic Infrastructures of Science in AI Era Should Rebalance Costs of Generation and Verification".
Jane: Here's the gist of the paper. AI systems can now cheaply generate plausible scientific artifacts such as papers, reviews, and surveys This creates a risk of epistemic pollution in our scientific systems where unreliable but plausible-looking artifacts can accumulate faster than the system can filter them The problem is structural as the epistemic infrastructure of science was calibrated to a world where producing a plausible artifact required substantial expertise, labor, and time so generation cost itself served as a rough filter This imbalance creates conditions for epistemic pollution the contamination of scientific literature by plausible yet unreliable artifacts Epistemic Pollution and Generation-Verification Imbalance The phenomenon is already visible across the surface area of the literature as large-scale analyses now find substantial LLM-associated changes in scientific writing and production patterns Examples include AI-generated surveys flooding preprint platforms termed a “survey paper DDoS attack” Peer review itself is under similar pressure with a study in Nature reported that 21% of manuscript reviews for a major AI conference were found to be generated by AI The diagnosis is therefore infrastructural rather than behavioral not that authors or reviewers have stopped doing their jobs but that the infrastructure they operate within was calibrated to a cost ratio that no longer holds Validity as the Verification Bottleneck Verification evaluates several dimensions of scientific claims including novelty significance and validity for two reasons Validity is often especially labor-intensive because reviewers can sometimes make preliminary judgments about novelty and significance from an abstract a figure and their knowledge of the field Second validity is the dimension where artifact design has the clearest direct leverage because it is fundamentally about the internal logic of an argument Validity verification often concentrates in validity and validity is the component whose cost depends most directly on how the artifact is internally organized Blueprints as Cost-Rebalancing Artifacts We propose blueprints as structured decomposed research artifacts designed to trade a generation cost of argument decomposition for cheaper downstream verification Blueprints decompose a project into entities such as claims evidence assumptions and definitions connected by relations such as supports expands and contradicts The prototype exposes two first-class interfaces respectively for human users and AI agents over a shared blueprint document The authoring workflow is similar in spirit to leanblueprint’s with adaptations forced by the move from formal mathematics to general research Examining the Verification Cost Verification cost is not a flat tax on the scientific record but a distribution across components paid many times by many parties Validity is different in two ways first it is often especially labor-intensive and second it is the component whose cost depends most directly on how the artifact is internally organized When an author writes a paper they hold something like a graph in mind which they compress into a linear text optimized for clarity persuasion and conformity to a venue’s expectations The diagnosis points to a specific design lever where the corresponding remedy is for the artifact to record that structure in the first place Blueprints Design and Usage A blueprint is a heterogeneous directed graph where nodes and edges are both typed The built-in vocabulary consists of seven node types organized into four roles five edge types and a small collection of node-type-specific status ladders The prototype exposes a browser canvas for human users and a command-line interface for AI agents over a shared blueprint document Structural lint surfaces what the reviewer should look at by checking the soundness of the support graph argument structure and vocabulary integrity The intended payoff is that review can become more local allowing a reviewer to inspect a single subgraph without first reconstructing the entire argument from prose AI has made plausible scientific artifacts cheap to generate while leaving reliable verification comparatively expensive We have suggested for one artifact-level response blueprints structured and decomposed argument graphs that make support relations inspectable before a verifier has to reconstruct them from prose Papers remain the narrative interface of science blueprints are proposed as a verification interface The broader argument is that AI-era scientific artifacts should be designed for verification as deliberately as they are designed for communication References
Paper summary: Jane: So we've talked about how this paper, "Epistemic Infrastructures of Science in AI Era Should Rebalance Costs of Generation and Verification," suggests that the problem isn't just one part—it’s the whole system being set up wrong for the age of cheap AI generation >
Tom: The core idea is that we need to treat this as an engineering challenge, redesigning how science artifacts are structured so that verification cost doesn't explode while generation cost plummets >
Lu: They propose moving away from prose-first narratives toward structured, decomposed representations of shorter local arguments, using something like leanblueprint to decompose a big theorem into smaller, verifiable claims >
Meng: From my side, the implication is that if we want AI-generated content to be useful and trustworthy in production environments, the artifact needs to be designed for verification as deliberately as it is designed for communication >
Lalam: The authors are essentially saying that AI-era scientific artifacts should be built with a focus on verifiability from the start, rather than just trying to communicate complex ideas in a single long piece of text >
Jane: So, to sum up, the paper argues that instead of letting the infrastructure get polluted by cheap but unreliable artifacts, we need to engineer new ways for science to be structured so that checking those artifacts becomes cheaper and more reliable >
Conclusion: Tom: So we’ve been looking at how cheap AI generation is getting for scientific papers, and this paper is about what that means for the whole system.
Jane: It’s called "Epistemic Infrastructures of Science in AI Era Should Rebalance Costs of Generation and Verification." The main idea they’re pushing is that the structure of science itself needs a redesign.
Tom: Exactly. They argue that right now, we’ve set up the whole system based on the idea that producing a good paper takes a lot of time and expertise. Now, if AI can make plausible-looking papers cheaply, it messes up those filters because they aren't calibrated anymore.
Jane: So they see this as an epistemic pollution problem. It’s like having junk pile up faster than anyone can sort it out because the system is built for a different cost ratio.
Tom: Right. And they point out that verification—actually checking if something is valid—is still super hard and expensive compared to just writing something down. That’s the core of their argument here.
Jane: They propose a solution called blueprints, which are like structured graphs for arguments instead of long prose narratives. These blueprints are designed to be cheaper to verify because they decompose the big argument into smaller, inspectable parts—like claims and evidence—before anyone tries to read it all at once.
Tom: So they’re suggesting that if we want AI-generated science to be useful, we have to design those artifacts for verification as much as we design them for communicating. It sounds like a fundamental shift in how the scientific community should even think about writing and checking work.
Jane: It really is about moving from just writing things down to designing arguments that are inherently verifiable. But what does this mean practically for the next big wave of AI research?
Jiaqi W. Ma
University of Illinois Urbana-Champaign
cs.CY, cs.AI
Submitted: 2026-05-11
Updated: 2026-11-10
Code: https://github.com/PatrickMassot/leanblueprint
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 100/100
The gist: AI systems can now cheaply generate plausible scientific artifacts such as papers, reviews, and surveys This creates a risk of epistemic pollution in our scientific systems where unreliable but…
Key concepts
- Epistemic Pollution
- This occurs when unreliable but plausible-looking scientific artifacts flood the literature faster than the system can filter them. Because AI generation is cheap, these misleading artifacts can contaminate our scientific knowledge base.
- Generation-Verification Imbalance
- The current scientific infrastructure was built assuming producing a plausible artifact required high expertise and time. Since generation cost is now very low, this imbalance means verification—checking for true validity—remains expensive and is the bottleneck.
- Blueprints
- These are structured, decomposed research artifacts that break down a project into components like claims, evidence, and assumptions connected by explicit relations (e.g., supports). They serve as a verification interface because they record the argument's internal structure directly.
Terminology
Summary
AI systems can now cheaply generate plausible scientific artifacts such as papers, reviews, and surveys This creates a risk of epistemic pollution in our scientific systems where unreliable but plausible-looking artifacts can accumulate faster than the system can filter them The problem is structural as the epistemic infrastructure of science was calibrated to a world where producing a plausible artifact required substantial expertise, labor, and time so generation cost itself served as a rough filter This imbalance creates conditions for epistemic pollution the contamination of scientific literature by plausible yet unreliable artifacts
Epistemic Pollution and Generation-Verification Imbalance
The phenomenon is already visible across the surface area of the literature as large-scale analyses now find substantial LLM-associated changes in scientific writing and production patterns Examples include AI-generated surveys flooding preprint platforms termed a “survey paper DDoS attack” Peer review itself is under similar pressure with a study in Nature reported that 21% of manuscript reviews for a major AI conference were found to be generated by AI The diagnosis is therefore infrastructural rather than behavioral not that authors or reviewers have stopped doing their jobs but that the infrastructure they operate within was calibrated to a cost ratio that no longer holds
Validity as the Verification Bottleneck
Verification evaluates several dimensions of scientific claims including novelty significance and validity for two reasons Validity is often especially labor-intensive because reviewers can sometimes make preliminary judgments about novelty and significance from an abstract a figure and their knowledge of the field Second validity is the dimension where artifact design has the clearest direct leverage because it is fundamentally about the internal logic of an argument Validity verification often concentrates in validity and validity is the component whose cost depends most directly on how the artifact is internally organized
Blueprints as Cost-Rebalancing Artifacts
We propose blueprints as structured decomposed research artifacts designed to trade a generation cost of argument decomposition for cheaper downstream verification Blueprints decompose a project into entities such as claims evidence assumptions and definitions connected by relations such as supports expands and contradicts The prototype exposes two first-class interfaces respectively for human users and AI agents over a shared blueprint document The authoring workflow is similar in spirit to leanblueprint’s with adaptations forced by the move from formal mathematics to general research
Examining the Verification Cost
Verification cost is not a flat tax on the scientific record but a distribution across components paid many times by many parties Validity is different in two ways first it is often especially labor-intensive and second it is the component whose cost depends most directly on how the artifact is internally organized When an author writes a paper they hold something like a graph in mind which they compress into a linear text optimized for clarity persuasion and conformity to a venue’s expectations The diagnosis points to a specific design lever where the corresponding remedy is for the artifact to record that structure in the first place
Blueprints Design and Usage
A blueprint is a heterogeneous directed graph where nodes and edges are both typed The built-in vocabulary consists of seven node types organized into four roles five edge types and a small collection of node-type-specific status ladders The prototype exposes a browser canvas for human users and a command-line interface for AI agents over a shared blueprint document Structural lint surfaces what the reviewer should look at by checking the soundness of the support graph argument structure and vocabulary integrity The intended payoff is that review can become more local allowing a reviewer to inspect a single subgraph without first reconstructing the entire argument from prose
Conclusion
AI has made plausible scientific artifacts cheap to generate while leaving reliable verification comparatively expensive We have suggested for one artifact-level response blueprints structured and decomposed argument graphs that make support relations inspectable before a verifier has to reconstruct them from prose Papers remain the narrative interface of science blueprints are proposed as a verification interface The broader argument is that AI-era scientific artifacts should be designed for verification as deliberately as they are designed for communication
References
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Redesign epistemic infrastructure to rebalance generation and verification costs by treating scientific artifacts as structured, decomposed representations called blueprints. This allows for an upfront generation cost of argument decomposition in exchange for cheaper, more local verification downstream, as stated:
Blueprints are designed to trade an upfront generation cost for cheaper, more local, more distributed verification downstream.
-
Implement a dual-interface system where a single blueprint document is accessible via both a browser canvas and a command-line interface. This enables human users and AI agents to edit the same artifact in parallel, allowing an AI agent to
inspect a finished argument
orextend a subgraph
as described in Section 4.3.1. -
Integrate structural linting into the blueprint system that flags specific misuse patterns, such as unsupported claims or circular support edges, to aid review without adjudicating content. This helps shift verification effort by surfacing what
the reviewer should look at,
allowing fordistributed and local review
where a reviewer can evaluate a single subgraph in parallel. -
Develop AI agents capable of performing task-specific verification on localized subgraphs rather than attempting to verify entire prose papers. An AI critic can be given
a single matured-claim subgraph and asked to enumerate plausible counter-arguments as new risk nodes,
directly addressing the bottleneck identified in Section 3.2 where validity verification is often especially labor-intensive due to prose compression. -
Design generation workflows that incentivize the explicit recording of argument structure over narrative flow, moving from prose-first narratives to graph-first representations. This ensures
the artifact records structure in the first place,
which is the key remedy proposed for validity verification cost concentration in Section 3.4.
Sources
- The Story is Not the Science: Execution-Grounded Evaluation of Mechanistic Interpretability Research
- Position: The AI Conference Peer Review Crisis Demands Author Feedback and Reviewer Rewards
- Stop DDoS Attacking the Research Community with AI-Generated Survey Papers
- The Last Human-Written Paper: Agent-Native Research Artifacts
- The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery
- HalluCitation Matters: Revealing the Impact of Hallucinated References with 300 Hallucinated Papers in ACL Conferences
- When AI Co-Scientists Fail: SPOT-a Benchmark for Automated Verification of Scientific Research
- AI-Researcher: Autonomous Scientific Innovation
- Position: The ML Community Must Build an AI-Augmented Peer-Review Ecosystem
- The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search
Related papers
- AI-Mediated Communication Can Steer Collective Opinion
- AI out-persuades expert humans in text-based dialogue
- Evaluating Ambient Clinical Scribes in India: The Need for Multilingual Real-World Clinical Conversation Data
- AI Value Alignment for Evolving Social Norms
- Voting Biases in Decentralized Autonomous Organization (DAO) Governance
- GenAIT: Development and Validation of an Objective Generative AI Literacy Test for High School Students