Systematic reviews in Elicit | Screening & extraction

Elicit’s PDF-based workflow is built to speed up the two most time-consuming parts of evidence synthesis—screening studies for inclusion and extracting structured data—while keeping a human-checkable audit trail. The core promise is that AI can pull key information out of papers (including tables), surface the most relevant supporting quotes, and flag low-confidence answers so reviewers can focus their time on verification and interpretation rather than copy-and-paste work.

The walkthrough centers on the “Extract data from PDFs” workflow. Users can upload multiple PDFs at once into their private library, then open each paper to view extracted text and tables. Elicit’s table extraction is highlighted as a differentiator: it can extract data from tables rather than limiting extraction to narrative text. For screening, the system supports column-based extraction aligned to review criteria. Reviewers can start with broad, open-ended fields such as “population characteristics,” then iteratively narrow the focus by adding more specific columns—like participant age or region—based on what the initial pass reveals.

A key operational feature is traceability. When Elicit fills a column, users can click into answers to see the underlying sources and the most relevant quotes, and they can open the paper to confirm context. When Elicit is uncertain, it marks the response with a confidence flag, prompting reviewers to double-check rather than accept everything at face value. The workflow also supports custom columns for criteria not covered by predefined fields. For example, instead of relying on a generic “region” field, a reviewer can create a “continent where the study took place” column with formatting instructions, then filter papers by keyword matches in that column (e.g., include Africa and Asia). This makes it possible to operationalize inclusion/exclusion rules directly inside the screening interface.

After screening and extraction, results can be exported as a CSV for use in spreadsheets like Excel. The export includes extracted metadata (title, authors, and the chosen columns), supporting quotes, and reasoning where applicable—often marking entries as “not applicable” or “not mentioned” with an explanation. Reviewers can then add their own tracking columns such as “reviewed” or “in progress,” using spreadsheet tools for sorting, filtering, and ongoing edits.

Accuracy controls matter. Elicit offers a “high accuracy mode,” described as reducing error by about half compared with regular mode, but at higher cost. The guidance is to use a rougher first pass for large-scale screening, then switch to high accuracy for later-stage data extraction—especially when detailed information is likely to live in tables (e.g., effect sizes).

Finally, the transcript cites internal testing against manual work. For screening roughly 5,000 papers, Elicit reportedly retrieved over 96% of studies deemed relevant by a team, compared with about 92% for trained human research assistants. For data extraction, the system is described as reaching around 98% accuracy versus about 72% for trained members in one comparison, with disagreement cases often resolving in Elicit’s favor after re-checking. The work is saved in the user’s sidebar for continuity, and the papers remain private to the user rather than shared publicly.