Counterfactual audit: LLM manuscript evaluators shift prestige and acceptance scores based on author bibliometrics and institution

Marco Rospocher, University of Verona; published in Discover Artificial Intelligence (Springer Nature) · Research evaluation and peer review (AI-assisted triage) · Italy-based study; 80 arXiv manuscripts; institutions from US/UK rankings

Research Universities

Publisher: Discover Artificial Intelligence (Springer Nature)
Original publication: Discover Artificial Intelligence, open access, 8 October 2026
Source retrieved: 2026-10-12

Read original source

What happened

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).

Why it matters

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 and measured results

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.

Limitations and uncertainty

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).

Put this evidence to work

Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-11; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.

Sales

Role takeaway

Customer problem: 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
Discovery questions: 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?
Value hypothesis: A counterfactual audit and redaction design may reduce metadata-driven shortlist changes before AI triage is used at scale; magnitude depends on model.
Potential engagement: Two-to-four-week metadata-sensitivity audit of one AI-assisted screening workflow.
Unsupported claims: LLMs are more biased than human reviewers
Effects generalize to all fields and languages
Redaction fully eliminates bias

Pre-sales engineering

Role takeaway

Fit: Institutions piloting LLM-based scoring for internal competitions, manuscripts or applications.
Architecture and integration: 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.
Proof of value: Metadata-swap test on 50+ past submissions
Measure top-K overlap with and without author fields
Re-run after any evaluator model update

Delivery

Role takeaway

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.
Skills and adoption: Train reviewers that AI scores are advisory and how to interpret audit results.
Governance checkpoints: Audit before first use
Equity review of results
Re-audit on model change
Acceptance criteria: 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

Implementation considerations

Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.

Architecture and integration

What must connect, and where does the AI sit in the workflow?

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.

Governance

Who approves, reviews and stays accountable for outcomes?

Institutions using LLMs in screening should audit for metadata sensitivity, re-audit on model changes and record metadata exposure.

Security and privacy

What data, permissions and controls need testing?

Retrieval-enabled assistants can look up author bibliometrics, creating a hidden metadata pathway; disable or log lookups.

Accessibility and workforce

Who is affected, and what skills or accommodations follow?

Prestige-sensitive AI screening could disadvantage early-career researchers and those at lower-ranked institutions.

Procurement

What should contracts, pricing and exit terms secure?

Vendors of AI review tools should be asked for metadata-invariance audits and evaluator-model change notices.

Operating model

Which teams own the service once it runs?

Keep final decisions with humans and separate content scoring from track-record assessment.

Publication history

2026-10-11Research Universities · Issue 342 resources

Read preserved resource versions (JSON)

Stable resource ID: verona-llm-peer-review-author-metadata-audit-20261008