What ASReview is
ASReview is an open-source tool that applies active learning to title-and-abstract screening in systematic reviews. You import a deduplicated set of records, label a few as relevant and a few as irrelevant, and a classifier continuously reorders everything you have not yet seen so that the records most likely to be relevant come next.
It comes from a research group at Utrecht University and was published in Nature Machine Intelligence in 2021, which is the fact that separates it from the rest of the AI-screening field. The method has a citable paper, an open codebase, and a simulation mode that lets you benchmark its behaviour against reviews whose answers are already known. When a methods reviewer asks what your screening tool actually does, you can answer with a reference rather than a vendor claim.
Why researchers use it
- The pile reorders itself — relevant records cluster at the front, so the tail of the screening list becomes increasingly empty of anything you need.
- Published, benchmarkable performance — simulation studies on completed reviews report finding the large majority of relevant records after screening a fraction of the total, and you can run those simulations yourself on your own data.
- Runs entirely on your machine — a local web interface, no upload, no account. Unpublished review data never leaves your laptop.
- Defensible in a methods section — open algorithms, versioned software, and a growing methodological literature on how to use and report it.
- Simulation mode — replay a review you have already completed to see how the model would have behaved before you trust it on a live one.
Where it fits in a research workflow
ASReview handles exactly one stage: title-and-abstract screening. Upstream sits your database search and deduplication, usually done in EndNote, Zotero, or a dedicated deduplication step, and exported as RIS or CSV. Downstream sits full-text review and data extraction, which ASReview does not do — that is Covidence territory, or a well-designed spreadsheet.
That narrow scope is deliberate and it is why ASReview slots into an existing protocol rather than replacing it. A common arrangement in a well-run review: search and deduplicate in your reference manager, prioritise with ASReview, screen collaboratively in Rayyan if two independent reviewers are required, then extract in Covidence. The tools do not compete; they cover different rows of your PRISMA diagram.
Getting started
An hour to install and understand, and do the simulation before the real thing.
- Install with
pip install asreviewand start the local interface withasreview lab. A Python environment manager such as uv or conda keeps this from polluting your system Python. - Run simulation mode first on a completed review — yours or one of the bundled benchmark datasets. This is how you learn what the ordering actually does before it is deciding your inclusions.
- Import your deduplicated RIS or CSV, then label roughly ten clearly relevant and ten clearly irrelevant records to prime the model.
- The step people skip, and the one reviewers will ask about: decide and write down your stopping rule before you start screening. “I stopped when it felt done” is not a method. Pre-register it in your protocol.
ASReview vs the alternatives
| Alternative | Does it better | Pick it if |
|---|---|---|
| Rayyan | Multi-reviewer blind screening and conflict resolution | Your protocol requires two independent screeners |
| Covidence | The whole pipeline — full text, risk of bias, extraction, PRISMA | You need one audited system from import to extraction table |
| Litmaps | Finding papers your search missed | You are still building the search, not screening its results |
| Semantic Scholar | Search and a free API for custom pipelines | You want to script discovery rather than prioritise screening |
ASReview is the only one of these built as a research method rather than a product. Use it for prioritisation; use the others for the surrounding workflow.
Cost, licensing, and your data
Free and open source under the Apache 2.0 licence, developed and maintained by an academic group with institutional funding rather than a commercial roadmap. There is no paid tier and nothing to negotiate.
Data governance is the quiet advantage. ASReview runs as a local server on your own machine; your records, your labels, and your in-progress decisions stay on disk. For reviews under embargo, for grant-competitive topics, and for any protocol where an unpublished search strategy is itself sensitive, this removes the entire question that cloud screening platforms force you to answer. It is also the reason it works on an air-gapped or institutionally locked-down machine where the browser-based alternatives simply cannot run.
The honest review
Strengths. ASReview is the rare AI research tool whose central claim is testable, and which hands you the means to test it. Simulation mode against known reviews turns “trust our model” into an experiment you run on your own corpus. That epistemic honesty — plus a peer-reviewed method paper and an open repository — is why this is the machine-assisted screening tool to cite.
Limitations. The stopping problem is genuinely unsolved and it is yours to own. Active learning tells you what to read next; it cannot tell you when the remaining pile is empty of relevant records, and the methodological literature on stopping criteria is still developing. Expect to justify your choice to reviewers, and expect some journals to be unpersuaded. Beyond that: the interface assumes you already know systematic-review practice, installation via pip puts a Python environment between you and the tool, and there is no multi-reviewer blind screening — this is a single-screener prioritisation engine, not a collaboration platform.
Verdict. Adopt it for any review where the screening set runs into the thousands and you have the methodological confidence to defend a stopping rule. Skip it for small reviews, for protocols that mandate exhaustive double screening of every record, and for teams that need collaborative workflow more than they need prioritisation. The condition that flips the answer is the size of the pile: below a few hundred records, setup costs more than it saves.
Common questions
Is ASReview free?
Yes, entirely. It is open source under the Apache 2.0 licence, developed at Utrecht University, with no paid tier, no account, and no usage limits. You install it yourself and run it locally.
How much screening time does ASReview actually save?
Published simulations on completed reviews report finding the large majority of relevant records after screening a fraction of the total, but the figure varies substantially with the review’s topic and how distinctive the relevant records are. Run simulation mode on a comparable dataset rather than assuming a headline number applies to you.
Can I use ASReview for a Cochrane review?
Prioritisation is compatible with many protocols, but stopping early on the basis of a model is a methodological decision that must be pre-registered and justified, and some review guidelines do not permit it. Confirm with your methodologist and your registration before relying on it to reduce screening volume.