New! Automate your Systematic Review with Elicit

Elicit’s new systematic review workflow aims to automate nearly the entire pipeline—refining the research question, collecting candidate papers, screening them against custom criteria, extracting structured data, and generating a first-draft report—while keeping every AI decision traceable to the underlying text. The practical payoff is speed without a “black box”: users can review why each paper was included or excluded, override judgments, and revise screening criteria or extraction fields before running on the full corpus.

The process starts with a research question. Users can begin broad (the example centers on microplastics and pregnancy), and Elicit provides hover-based suggestions that tighten key assumptions—such as specifying the population, exposure measurement approach, and outcome definition. Once the question is selected, the workflow can auto-populate the next steps with AI-generated recommendations, though users can uncheck automation to work manually.

Paper collection happens in a “Gather” step with three main routes. First, Elicit Search can pull up to 500 papers at once using semantic search over more than 126 million papers from the Semantic Scholar database, ranking relevance based on title and abstract meaning rather than keyword matching alone. Second, users can add papers from their Elicit Library using tags. Third, users can upload PDFs directly. Elicit positions this as complementary to traditional Boolean or database-specific searching (including workflows still common for regulatory or publication standards): keyword search can remain the baseline, while semantic search helps reduce the risk of missing relevant studies.

Next comes screening criteria definition. Elicit generates custom screening columns tailored to the research question—questions about exposure assessment (e.g., direct maternal exposure vs environmental levels), quantitative measurement, participant type (humans vs animals), study design (excluding case reports, for instance), and whether birth outcomes are reported. Users can toggle suggested criteria on or off, add their own, and iterate quickly using a representative sample of 100 papers before committing to the full set. Each screening decision is presented with explicit yes/no/maybe judgments plus explanations tied to the paper’s text, and papers can be ranked by likelihood of inclusion for faster review. Importantly, screening is not treated as one-and-done: criteria can be adjusted later, and overrides are allowed at the paper level.

After screening, extraction follows a similar pattern: Elicit auto-suggests extraction fields in a “PICO-like” structure (customized to the question) and provides detailed instructions for what to pull from methods and results. The system supports both free-form responses and structured options such as yes/no/maybe, and it can extract quantitative data from tables. Transparency is built in via one-click links to quotes from the source text (including table citations) whenever AI fills a field.

Finally, Elicit generates a report designed as a usable first draft rather than a final publication-ready synthesis. The report includes a one-line summary, a PRISMA-inspired diagram of how studies were evaluated across screening criteria, summaries of included studies, and thematic discussion. Users can chat with the report and underlying data, export results at each stage, download the report as a PDF, and collaborate in real time on supported plans. The workflow is also framed as enabling “living reviews,” where adding new papers can flow through existing criteria and fields rather than restarting from scratch.