AI tools for meta-analysis: what they can do at each stage
A meta-analysis is not one task but about eight: choosing a question, searching, screening, extracting data, judging risk of bias, pooling the results, grading the evidence and writing it up. AI tools exist for most of them, and they are not equally trustworthy. Ranking records for screening is well studied. Asking a chat assistant for a pooled estimate is not something you can defend in a methods section.
This guide goes through the stages in order, says what AI can reasonably do at each one and what to check before you rely on it, and explains why the statistics themselves belong in validated software. SynthGap is our own tool and covers only the first stage, so read this as a guide from a vendor, not as an independent comparison.
The short answer
Treat AI as an assistant at specific steps, not as the author of the review. Several tools can save real time on parts of a meta-analysis, as long as a person checks what they produce and the use is reported.
Cochrane, the Campbell Collaboration, JBI and the Collaboration for Environmental Evidence took that position together in 2025. Synthesists stay responsible for their review and may use AI if they can show it does not compromise rigour, and they are expected to keep human oversight and to report any use that makes or suggests judgements. How to report AI use in a systematic review shows what that looks like in practice.
Research groups are also building pipelines that try to run a whole review automatically. They are worth watching, but a methods paper is not the same as a validated tool that you can cite in your own protocol.
Stage by stage
The last column names examples of the kind of tool, not recommendations. We have not tested them for you, and features and terms change, so check each tool's own site.
| Stage | What AI can do | What to check | Examples |
|---|---|---|---|
| 1. Topic and question | Scan a field for gaps, propose candidate questions and show which already have reviews. | Open every study it cites and run your own search for existing reviews. A topic can look new because the tool missed a source. | SynthGap; AI research assistants |
| 2. Search | Draft Boolean strings, suggest synonyms and subject headings, translate a string between databases. | Subject headings that do not exist, results that change between runs, and lower recall than an expert-built search. | Chat assistants, with a librarian's check |
| 3. Screening | Rank records by likely relevance while you screen, or act as a second screener. | Recall on a sample you screened by hand, a stopping rule fixed in advance, and a log of what the tool excluded. | ASReview, Rayyan, Covidence, EPPI-Reviewer |
| 4. Data extraction | Pull study details and numbers from PDFs into a table. | Every number against the paper: standard deviation or standard error, the right arm, the right time point. | Elicit, SciSpace, chat assistants |
| 5. Risk of bias | Suggest a rating for each domain and the sentence in the paper that supports it. | Treat it as a second opinion. A person makes and records each judgement. | RobotReviewer (randomised trials) |
| 6. Pooling | Explain the methods and draft code for you to run. | Never use a pooled result that a language model worked out itself. Run the analysis in validated software and keep the script. | RevMan, R (metafor, meta), Stata, JASP |
| 7. Certainty of evidence | Draft a GRADE summary table for a person to confirm. | The ratings are judgements and need someone with content and methods knowledge. | Chat assistants (drafts only) |
| 8. Writing and reporting | Edit language, draft plain-language text, check a draft against the PRISMA 2020 checklist. | Every statement of fact against your own data, every reference against its source, and a declaration where AI made or suggested a judgement. | Chat assistants |
Let SynthGap run the first pass
Enter an area. SynthGap proposes candidate topics, checks each against existing systematic reviews, and lists the primary studies behind it. Start on the Free plan: no card needed, and every run shows every topic it finds.
Sign up freeWhere the statistics belong
The pooled estimate, the heterogeneity statistics and the forest and funnel plots should come from meta-analysis software that is validated and that you can run again, fed with numbers that a person has checked against the papers. Meta-analysis software: R, RevMan, Stata and more compares the options.
A language model can still help here. It can explain what an output means, help you choose between a fixed-effect and a random-effects model, or write R code. But if you ask it to calculate a pooled odds ratio and it answers in prose, you cannot tell whether any calculation took place. If an assistant does run code, ask for the code and the data it used, and run them yourself.
Where SynthGap fits
SynthGap works before the review starts. You describe a clinical area; it searches the published literature, checks candidate topics against existing systematic reviews, and lists the primary studies behind each topic so that you can open and check them. A deep analysis then reads the studies for a topic you choose and writes a critical report on it. Using AI to find a systematic review topic covers that stage in more detail.
It does not screen your search results, extract your data or pool your results. Its suggestions are AI-generated, so verify them before you register a protocol. In the table above it belongs to the first row, and it also helps you count the primary studies when you are planning a meta-analysis.
Questions to ask any tool
Four to start with. How to choose an AI tool for a systematic review has the full list of twelve.
- Evidence: has it been tested on reviews like yours? Ask for evaluations, ideally by people who do not sell the tool.
- Sources: can you see where each claim comes from? Look for a link to the study, not just a summary.
- Records: can you export what it did? You need the decisions, the settings and the version to report them.
- Data: where does your text go? Check the terms before you upload full texts or an unpublished protocol.
Sources
The figures and rules on this page come from these sources.
- Flemyng E, Noel-Storr A, Macura B, et al. Position statement on artificial intelligence (AI) use in evidence synthesis across Cochrane, the Campbell Collaboration, JBI and the Collaboration for Environmental Evidence 2025. Campbell Systematic Reviews 2025;21:e70074.
- Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 2021;372:n71.
Questions
Can AI do a whole meta-analysis for me?
Not in a way that meets the standards the evidence-synthesis organisations set. AI can speed up parts of it, such as ranking records for screening or drafting analysis code, but the judgements and the pooled result need a person and validated software. Pipelines that attempt every step exist as research projects, so check any claim of a fully automated review against the evidence for each step.
What is the best AI tool for meta-analysis?
There is no single best tool, because the stages differ. For finding and testing a topic, SynthGap is built for that, and it is our own tool. For screening, look for tools with published evaluations. For the statistics, use validated meta-analysis software and use AI only to explain output or draft code. Whatever you pick, pilot it on a sample you can check.
Is it acceptable to use AI in a meta-analysis I want to publish?
Often yes, if you can justify it, keep a person in charge of every judgement and report it. The PRISMA 2020 checklist asks for details of automation tools used in screening (item 8), data collection (item 9) and risk of bias assessment (item 11), and the journal's own AI policy applies as well. Check the policy of your target journal before you start.
Does AI make a meta-analysis faster?
It can shorten some steps, mainly screening, but the saving depends on how many records your search returns and on how much checking you do. Most of the remaining time goes on careful work that AI does not replace: eligibility decisions, extraction checks and the analysis plan.
Keep reading
General methodological guidance, not medical advice. SynthGap's suggestions are AI-generated from the published literature, so verify them before registering a protocol.