Qualitative Analysis of Discrepancy Patterns Between Large Language Models and Human Reviewers in Abstract Screening for Systematic Reviews.

Background

We evaluated the reliability of large language models (LLMs) for abstract screening under real-world review practices and qualitatively characterized model-human discordance to inform safe workflow integration.

Methods

We evaluated GPT-4.0, GPT-5.0, and GPT-5.0-mini on two curated systematic review datasets representing contrasting topic densities, defined by the target-to-background ratio (TBR): a core-subject dataset (TBR 69%) in which the target intervention was central, and a peripheral-subject dataset (TBR 4.4%) in which the target intervention was incidental. Final inclusion after full-text review served as the reference standard. We developed a qualitative taxonomy of disagreements, classifying false negatives as intended human leniency, gray-zone ambiguity, or true LLM misses, and false positives as implicit or additional human exclusion rules, gray-zone ambiguity, or nominal inclusions that increase workload only.

Results

GPT-5.0-mini achieved the best sensitivity-efficiency trade-off (core-subject: 91% sensitivity with 96.7% workload reduction; peripheral-subject: 83% sensitivity with 92.7% workload reduction) and negative predictive value >99% in both datasets. Disagreement was lower when relevance was central (core-subject: 1.6%, 7/430) with no true LLM misses (0/430). In the peripheral-subject dataset, disagreement was higher (10.6%, 74/696), driven mainly by intended human leniency among false negatives (52/56) and gray-zone ambiguity among false positives (12/18), while true LLM misses remained rare (0.4%, 3/696).

Conclusion

Many model-human disagreements reflect topic- and workflow-dependent screening conventions rather than intrinsic model failure. LLM-assisted screening may improve efficiency without compromising reliability when accompanied by appropriate safeguards for ambiguous records.