A Critical Review of AI Literature-Screening Tools: Elicit, ResearchRabbit, and Covidence

The most useful comparison is based on the stage of the review at which each platform adds value. During early exploration, Elicit can help locate and assess papers against a question. When the research landscape is fragmented, ResearchRabbit can reveal connections that keyword searching alone may overlook. Once protocol and eligibility criteria have been established, Covidence can help reviewers apply those criteria, manage disagreements and document decisions.

Speed should never be the only measure of success. In screening, sensitivity is especially important because it reflects how well a method retains studies that are genuinely eligible. A tool that excludes irrelevant records quickly but also removes relevant studies can distort the evidence base. Dennstädt et al. (2024) examined large language models for biomedical title-and-abstract screening and showed that performance varied with the model, dataset and decision threshold. Li et al. (2024) likewise found that prompt design and characteristics of the review dataset influenced the performance of ChatGPT-based screening. Specificity and precision also matter because they indicate how effectively a screening approach avoids retaining ineligible records; the appropriate balance depends on the review’s tolerance for missed studies and unnecessary full-text assessment.

Sensitivity is the proportion of all eligible studies that the screening approach correctly retains; specificity is the proportion of ineligible studies that it correctly rejects. Precision is the proportion of records classified as eligible that are genuinely eligible. A false exclusion is an eligible study incorrectly rejected, whereas a false inclusion is an ineligible study incorrectly retained. In a systematic review, false exclusions are especially serious because they can remove relevant evidence from the synthesis; however, low specificity or precision can create substantial unnecessary full-text assessment. These measures should be interpreted against a human-verified reference set and the review’s stated priorities, rather than assumed from a tool’s feature list.

The platforms also differ in their support for collaboration and traceability. Elicit can provide source passages behind some screening and extraction outputs; Covidence provides a structured record of reviewer decisions and disagreements; ResearchRabbit makes literature relationships easier to explore visually. None of these features alone establishes methodological validity. A credible comparison would test the tools on the same set of references, using the same inclusion criteria and decisions made independently by qualified reviewers as the reference standard. This article compares the platforms’ stated roles and workflow fit; it does not report an independent head-to-head performance test of Elicit, ResearchRabbit, and Covidence.

The practical trade-off is therefore one of function rather than overall superiority. Elicit brings search, screening and extraction functions into a guided workflow, but feature access and review controls vary by plan. ResearchRabbit is useful for citation-led expansion and visual exploration, but citation connections do not establish eligibility and may not provide comprehensive database coverage. Covidence offers a more structured environment for collaborative screening and decision tracking, but it does not design the research question or replace a reproducible database search. Selection should be based on the review protocol, the stage requiring support, team size and the level of auditability needed.