LitRev uses two independent AI models — GPT-4o (OpenAI) and Claude (Anthropic) — to screen every title, abstract, and full text against your inclusion criteria.
Why two models
A single AI model can be wrong in systematic ways. Two models trained independently are unlikely to make the same mistakes. When both agree, the decision is high-confidence. When they disagree, it's flagged for your review.
This mirrors the gold-standard "two independent reviewers" methodology from Cochrane and PRISMA 2020 — except your second reviewer is also you, with AI doing the first pass.
Confidence levels
Every AI decision comes with a confidence rating:
- HIGH — Both models agree, with strong evidence in the abstract or full text
- MODERATE — Both models agree, but evidence is partial or could be interpreted multiple ways
- LOW — Both models agree but with weak signals
- UNCLEAR — Models disagree; needs your judgment
LitRev's "Accept All AI" buttons only apply to HIGH-confidence decisions by default. LOW and UNCLEAR cases always stay for manual review.
What the AI sees
For title/abstract screening: the AI reads the title, abstract, and any keywords you've imported.
For full-text screening: the AI reads up to the first 20,000 characters of the PDF — typically covering the abstract, introduction, methods, results, and part of the discussion. This is enough for inclusion/exclusion decisions in over 95% of cases.
Auditability
Every AI decision is logged with:
- The exact reasoning the model gave
- The specific quotes from the text that justify the decision
- A primary PRISMA exclusion category (Wrong population, Wrong intervention, etc.)
When you export your PRISMA flow, this categorization automatically produces the required "reasons for exclusion" breakdown.
You make the final call
The AI is a fast first-pass assistant. You override any decision with one click. Your final included list — and your published systematic review — is your decision, not the AI's.