{"resourceId":"verona-llm-peer-review-author-metadata-audit-20261008","versions":[{"version":"external-cd1b46d8273c983660efb03f22a3e1b62cb108a33fac8fe453a3825dd20c8568","resource":{"id":"verona-llm-peer-review-author-metadata-audit-20261008","title":"Counterfactual audit: LLM manuscript evaluators shift prestige and acceptance scores based on author bibliometrics and institution","organization":"Marco Rospocher, University of Verona; published in Discover Artificial Intelligence (Springer Nature)","sector":"Research evaluation and peer review (AI-assisted triage)","geography":"Italy-based study; 80 arXiv manuscripts; institutions from US/UK rankings","publishedAt":"Discover Artificial Intelligence, open access, 8 October 2026","publicationDate":"2026-10-08","eventDate":null,"sourceName":"Discover Artificial Intelligence (Springer Nature)","sourceLabel":"Peer-reviewed academic study","sourceUrl":"https://link.springer.com/article/10.1007/s44163-026-02268-y","evidenceClass":"academic-research","outcomeClass":"cautionary","topics":["governance-procurement","knowledge-work","operating-model"],"finding":"A controlled 2x2x2 counterfactual audit of 80 recent arXiv manuscripts (CS, math, physics, q-bio) with six LLM evaluators, five repetitions each, found outputs are not invariant to corresponding-author metadata. High bibliometrics (h-index 32 vs 8) and high-tier institutions raised prestige/impact and acceptance judgments; content-quality effects and name-coded effects were smaller. For acceptance, 56 manuscripts scored higher under the high-bibliometric profile, 15 lower, 9 unchanged (FDR-adjusted p<0.001). Bibliometric perturbations removed roughly one quarter to nearly two fifths of top-K manuscripts. Single-evaluator agreement was low (ICC(A,1) 0.112–0.213).","sledRelevance":"Directly relevant to universities and sponsored-research offices considering LLMs for proposal screening, internal competitions, hiring or award triage: author-context metadata can change shortlists. Supports content-only, decoupled pipeline designs.","evidence":"Opened the full open-access article on Springer Nature Link (HTTPS fetch, full text read including methods, results, limitations). Also read a Scienmag summary for context.","architectureImplications":"Interpretation: evaluation pipelines should redact author metadata before LLM scoring and log which fields the model saw, keeping any track-record input in a separate module.","governanceImplications":"Institutions using LLMs in screening should audit for metadata sensitivity, re-audit on model changes and record metadata exposure.","securityPrivacyImplications":"Retrieval-enabled assistants can look up author bibliometrics, creating a hidden metadata pathway; disable or log lookups.","caveats":"No human-reviewer baseline; 80 English manuscripts in four hard-science domains; mostly small open models plus gpt-4.1-mini; single prompt template; explicit metadata in prompt rather than retrieved; author is on the journal's editorial board (declared, not involved in handling).","streamIds":["research"],"roles":{"sales":"Sales framing: research offices exploring AI-assisted proposal or manuscript screening face a documented fairness risk, since author bibliometrics and institution can change LLM shortlists. Ask whether author fields reach the model and whether outputs are audited. A bounded audit of an existing or planned screening workflow fits. The evidence cannot support claims that LLMs are more biased than humans, since there was no human baseline.","engineering":"Engineering framing: the mitigation is architectural. Redact author identifiers and affiliation before LLM scoring, disable lookups, and log the fields each evaluation saw. Constraints include model-specific effect sizes and prompt dependence. Validate with a counterfactual test: score the same submissions with and without swapped author profiles and measure shortlist changes before deployment.","delivery":"Delivery framing: deploying AI triage responsibly means building an audit into the process, not a one-time check. Work includes redaction, logging, a metadata-swap test set and re-audit triggers on model updates. Ownership sits with the research office and IT governance. Propose acceptance measures such as shortlist overlap above a set threshold under metadata swaps; risks include disparate impact on early-career applicants."},"roleDetails":{"sales":{"customerProblem":"Research offices want AI help screening proposals or manuscripts without importing prestige bias into shortlists.","stakeholders":["Vice president for research","Sponsored programs office","Faculty governance/equity office","IT and AI governance"],"discoveryQuestions":["Do author names, affiliations or metrics reach the model?","Can the tool retrieve author information on its own?","Has the workflow been tested with swapped author profiles?"],"valueHypothesis":"A counterfactual audit and redaction design may reduce metadata-driven shortlist changes before AI triage is used at scale; magnitude depends on model.","potentialEngagement":"Two-to-four-week metadata-sensitivity audit of one AI-assisted screening workflow.","unsupportedClaims":["LLMs are more biased than human reviewers","Effects generalize to all fields and languages","Redaction fully eliminates bias"]},"engineering":{"fit":"Institutions piloting LLM-based scoring for internal competitions, manuscripts or applications.","architectureIntegration":"Redaction step before LLM scoring, retrieval disabled, logging of metadata exposure, separate track-record module (interpretation drawn from the paper's decoupling proposal).","prerequisites":["Access to scoring pipeline","Sample of past submissions","Ability to disable model retrieval"],"constraints":["Effect sizes vary by model","Single prompt template studied","Hard-science English corpus"],"security":"Disable or log any web or database lookups by evaluator models to close hidden metadata pathways.","proofOfValue":["Metadata-swap test on 50+ past submissions","Measure top-K overlap with and without author fields","Re-run after any evaluator model update"]},"delivery":{"work":"Implement redaction, exposure logging, a counterfactual test set and re-audit triggers in AI-assisted screening.","dependencies":["Vendor or in-house pipeline access","Historical submissions","Governance approval"],"ownership":"Research office with IT/AI governance.","skillsAdoption":"Train reviewers that AI scores are advisory and how to interpret audit results.","governanceCheckpoints":["Audit before first use","Equity review of results","Re-audit on model change"],"acceptanceCriteria":["Proposed: top-K overlap of at least 95% under metadata swaps","Proposed: metadata exposure logged for 100% of evaluations","Proposed: human makes every final decision"],"risks":["Disparate impact on early-career or lower-tier applicants","Model drift","False reassurance from limited test sets"]}},"retrievedAt":"2026-10-11T21:58:00-05:00","enrichedAt":"2026-10-11T21:59:53-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Prestige-sensitive AI screening could disadvantage early-career researchers and those at lower-ranked institutions.","procurementImplications":"Vendors of AI review tools should be asked for metadata-invariance audits and evaluator-model change notices.","operatingModelImplications":"Keep final decisions with humans and separate content scoring from track-record assessment.","sourceVerification":{"openedUrl":"https://link.springer.com/article/10.1007/s44163-026-02268-y","referenceExcerpt":"Evaluator outputs are not invariant to author metadata.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}