AI evidence extraction

Scientific autonomy in the structural bubble: from institutional bias to AI-mediated consensus

Elias Rubenstein · Frontiers in Research Metrics and Analytics · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

3/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.3389/frma.2026.1766504

Methodology & findings

Study design

Conceptual synthesis integrating meta-research findings on conflicts of interest, peer review, evaluation systems, consensus infrastructures, and AI-mediated workflows into a unified multi-layer theoretical model.

Primary method

No empirical statistical methods were employed. The paper proposes a theoretical framework and derives testable hypotheses without conducting statistical analysis.

Main result

The manuscript identifies that "AI-mediated access can shift epistemic authority from evidence-weighted appraisal toward prestige-weighted surfacing unless constrained by traceable provenance and validation-loop governance." Under weak safeguards, AI systems "can intensify it through inherited bias, citation concentration, and over-generalization" of upstream structural dependencies in science, while "under strong workflow safeguards (traceability, validation loops, and oversight), those effects should weaken and can partially reverse."

Research paradigm

Critical realist; science and technology studies (STS)

Author conclusions

The authors conclude that "Large language models do not dissolve this structure. Under weak safeguards, they can intensify it through inherited bias, citation concentration, and over-generalization. Under strong workflow safeguards (traceability, validation loops, and oversight), those effects should weaken and can partially reverse. The proposed empirical program provides a direct route to testing these conditional claims." They emphasize that "scientific authorship must be understood primarily as the capacity to formulate original questions, to identify neglected lines of inquiry, to develop conceptual frameworks that make hidden assumptions explicit, and to take responsibility for epistemic and ethical implications."

Risk of bias

Publication bias in systematic filtering of research agendas; Prestige-weighted retrieval biases in AI systems; Funding source alignment with research agendas; Journal hierarchy effects on visibility and authority; Consensus-substrate lock-in effects (Wikipedia dominance); Automation bias in editorial workflows; Citation concentration and over-representation of high-prestige venues; Selection bias in funding allocation toward agenda-aligned projects; Publication bias favoring high-impact venues and prestige-weighted sources; Peer review bias documented across multiple studies (Wennerås and Wold 1997; Tomkins et al. 2017); Inherited bias in LLM training data reflecting historical publishing patterns; Citation bias in LLM-generated references toward already highly-cited papers; Automation bias in editorial and gatekeeping workflows; Metrics performativity bias where bibliometric indicators reshape what counts as legitimate contribution; Prestige-weighted retrieval in AI systems amplifying existing publication hierarchy biases; Consensus substrate lock-in through inherited training data biases; Citation bias toward already highly cited papers and prestigious venues; Automation bias in editorial and review decision-making; Language and regional representation biases in LLM training corpora; Over-generalization in AI-mediated synthesis compressing boundary conditions and limitations

Limitations

The paper acknowledges that its mechanistic sketch "is not a microfounded causal model of scientific authority, and it does not identify causal effects without additional assumptions." The authors note that "in empirical use, A(s), E(s), and R(s) are proxies that require field-specific operationalization and audit
functional forms need not be linear or stable across fields." Additionally, the treatment of Wikipedia as representative of consensus infrastructures is noted as illustrative rather than exhaustive: "Wikipedia is used here primarily as an illustrative case of a broader class of policy-governed secondary synthesis and reference platforms
the mechanism proposed in this manuscript does not depend on Wikipedia specifically."

Open questions raised

Empirical testing of upstream topic selection effects under structural dependence (Hypothesis H1)
Measurement of visibility consolidation via publication and evaluation systems (Hypothesis H2)
Testing of consensus lock-in mechanisms under AI mediation with weak safeguards (Hypothesis H3)
Natural experiments examining AI-assisted gatekeeping effects on acceptance patterns
Domain-specific operationalization of autonomy proxies (agenda diversity, attribution diversity, long-tail inclusion)
Empirical testing of consensus lock-in effects under AI-mediated access

Data: None. This is a conceptual synthesis article without primary data collection.Code: None mentionedExtracted from: pdfAgreement 65%

Explore related topics

Epistemic Risk AI Governance Conceptual