Can a Large Language Model Serve as the Missing Second Reviewer? Prominence Bias, Retrieval Effects, and a Confidence-Fabrication Gap in an Empirical Evaluation of Two Published Meta-Analyses

Technical Summary: Can a Large Language Model Serve as the Missing Second Reviewer?

Problem Statement
Single-author systematic reviews (SRs) and meta-analyses (MAs) are prevalent in biomedical literature, despite methodological guidance (e.g., the Cochrane Handbook) discouraging single authorship due to the risks of missed studies, transcription errors, and selection bias. The standard mitigation is independent dual screening and extraction by a second human reviewer. Large Language Models (LLMs) have been proposed as a mechanism to supply this "missing second reviewer" without requiring a second human collaborator. However, existing validation studies are limited by small sample sizes (4–26 reviews), single-condition testing, and a lack of mechanistic characterization regarding how LLMs succeed or fail in these specific tasks.

Methodology
This study presents an empirical evaluation of a current-generation, publicly available LLM (Claude, Anthropic) acting as a second reviewer across two distinct task designs, tested on two published meta-analyses selected for topical and temporal separation:

1. Design B (Second-Reviewer Extraction Check): The model was provided with a known list of nine included Randomized Controlled Trials (RCTs) from a metformin-and-cancer meta-analysis. The task was to independently extract outcome measures, effect sizes, and confidence intervals, flag data-quality concerns, and calculate a pooled estimate. The model was tested across 8 independent runs using two prompt-schema iterations (v1 and v2), with the latter adding structured fields for "sourcing_basis" and "data_quality_flag."
2. Design A (Autonomous Replication): The model was provided only with a PICO research question and a literature cutoff date for a multivessel-versus-culprit-only PCI meta-analysis. The task was to independently identify included studies, extract data, and synthesize a conclusion. This design was tested in two conditions:
   • Without Retrieval: 3 runs relying solely on parametric training knowledge.
   • With Retrieval: 5 runs explicitly instructed to use web search tools, with verbatim query logging and disclosure of search usage.

All model outputs were independently verified by the author against primary sources (original trial reports) to distinguish genuine errors from model fabrications.

Key Results

• Error Detection and Fabrication (Design B):

  • The model consistently identified an internal anomaly in the source meta-analysis: a duplicated hazard ratio and confidence interval reported for two distinct endpoints in one trial.
  • Across 8 runs, the model flagged four additional discrepancies confirmed as real errors in the published meta-analysis (e.g., sample-size transcription errors, blinding inconsistencies, effect-size mismatches) that had been missed by the original two-person human review team.
  • Crucially, the model also produced one confirmed fabrication: an invented internal numerical inconsistency for a trial that did not exist in the source.
  • Confidence Gap: The model's confidence language and self-reported "sourcing_basis" did not reliably distinguish between genuine catches and fabrications. A fabrication was delivered with the same structured labeling and tone as genuine findings.

• Study Identification and Retrieval (Design A):

  • Without Retrieval: Mean recall of the actual included trials was 55.6%. A clear prominence bias was observed: the model correctly identified prominent, guideline-referenced trials but systematically missed smaller, earlier, or narrower-scope trials.
  • With Retrieval: Mean recall improved significantly to 87.5%. However, retrieval did not eliminate precision errors: the model frequently conflated studies, importing real, topically related trials that were not part of the specific paper's reference list (cross-paper conflation).
  • Synthesis Failure: In neither condition (with or without retrieval) did the model achieve independent quantitative pooling. Runs either cited existing pooled estimates from other meta-analyses or calculated partial pools from a subset of studies, explicitly flagging the rest as unpooled due to heterogeneity or missing raw data.

Key Contributions

1. Mechanistic Characterization: The study moves beyond aggregate accuracy metrics to characterize specific failure modes: the tendency to fabricate inconsistencies with high confidence, the "prominence bias" in autonomous study identification, and the inability of retrieval tools to resolve paper-specific inclusion precision.
2. Validation of Supplementary Utility: The results demonstrate that an LLM can catch real, independently verifiable errors that a multi-person human review team missed, suggesting value as a supplementary check rather than a replacement.
3. Schema and Protocol Improvements: The study highlights the necessity of structured disclosure fields (e.g., "sourcing_basis") and the insufficiency of self-reported confidence. It establishes that retrieval improves recall but does not solve provenance verification or independent computation.

Significance and Claims
The paper modestly concludes that LLM-assisted second review is complementary, not superior, to human verification.

• Support for Complementary Framing: The data supports deploying LLMs to surface real errors and improve study-identification recall, particularly when equipped with retrieval tools and structured disclosure requirements.
• Rejection of Substitution: The findings do not support using LLMs as a substitute for human verification. The presence of confident fabrications and persistent precision/synthesis limitations means that every LLM output still requires independent verification—the same requirement that motivates dual human review in the first place.
• Design Principles: The study proposes three concrete design principles for responsible deployment: (1) mandatory structured sourcing and data-quality disclosure fields; (2) retrieval-enabled default conditions with primary-source text supplied for extraction; and (3) scoring per-outcome pooled effects rather than relying on a single collapsed estimate.

The authors note limitations, including the small scale (two meta-analyses), the use of citation-only inputs in Design B (rather than primary source text), and single-author verification, framing these findings as a standalone contribution intended to inform future, larger-scale evaluations.