What Elicit is
Elicit is an AI research assistant purpose-built for literature work. You ask a research question in plain language; it retrieves real papers from a large academic corpus, and — the part that distinguishes it — extracts structured fields from each one into a table you can compare across: sample size, population, intervention, outcome measure, limitations.
The design decision that matters is that Elicit retrieves before it generates. It searches an actual index of published papers and then summarises what it found, rather than producing text about a topic and attaching references afterwards. That ordering is why it does not hallucinate citations the way a general assistant does — the papers exist because they came out of a database. It does not make the extraction infallible, which is the subject of the limitations below, but it removes the single worst failure mode in AI-assisted literature work.
Why researchers use it
- The extraction table — this is the product. Twenty papers, five columns, one screen; the comparison you would otherwise build by hand over two days.
- Question-shaped search — ask “does X improve Y in older adults” rather than constructing Boolean syntax before you know the field’s vocabulary.
- Real papers with links — every row traces to a paper you can open, which makes verification a click rather than an investigation.
- Custom columns — define the field you actually care about and have it extracted across your whole set, which is the closest thing to automated data extraction that exists for scoping work.
- Export that goes somewhere — BibTeX and RIS out to Zotero, CSV out to your analysis.
Where it fits in a research workflow
Elicit occupies the gap between finding papers and extracting from them, which is exactly where most literature reviews lose their weeks. Upstream is a question and a rough sense of the field; downstream is either a written synthesis or a formal review pipeline.
For a scoping review or a thesis introduction, it can be most of the work — search, screen, extract, export to Zotero, write. For a registered systematic review it is an accelerator inside a protocol you still have to run properly: your reported search belongs in a named database, screening belongs in Rayyan or ASReview, and formal extraction belongs in Covidence with two independent reviewers. Elicit’s extraction is a first pass that a human checks, not a reviewer. If you want to build this pattern yourself for a specific field, the structured-outputs extraction workflow shows the same idea in fifty lines of Python.
Getting started
Half an hour, and the fourth step is what makes it trustworthy.
- Ask one specific, answerable question about a relationship between named things. Broad topics return broad noise here as everywhere.
- Skim the returned papers, star the relevant ones, and build a working set of ten to twenty before extracting anything.
- Add the columns you actually need — not the defaults. “Follow-up duration” or “control condition” is worth more to you than a generic summary.
- The step people skip, and the one that decides whether this tool helps or hurts: open three papers and check the extracted values against the actual text. Do it early. You will learn exactly which fields it gets right and which need eyes, and that calibration is what makes the other seventeen rows usable.
Elicit vs the alternatives
| Alternative | Does it better | Pick it if |
|---|---|---|
| Consensus | Fast orientation on whether a claim is contested | You want a fifteen-minute read of the field, not a table |
| Semantic Scholar | Free search and an open API for custom pipelines | You would rather build the extraction yourself |
| ASReview | Peer-reviewed screening prioritisation for thousands of records | Your bottleneck is screening volume, not extraction |
| Covidence | Auditable dual extraction with reconciliation | The review is registered and must satisfy a methodologist |
Elicit is the strongest of these at the specific job of turning a pile of papers into a comparable table. It is not a substitute for any of the others at their own jobs.
Cost, licensing, and your data
Freemium: a free tier with a monthly allowance of credits covering searches and extractions, paid individual plans for heavier use, and team and institutional options. The credit model means cost scales with how much you extract rather than how long you use it, which suits bursty literature-review work well.
Nothing sensitive is involved in searching published literature. Two things are worth naming anyway. First, your questions and custom extraction columns describe your research direction in unusual detail to a third party — normally fine, occasionally not, for a competitive grant. Second, and more practically: uploading your own PDFs for extraction means sending publisher-copyright files to a service, which some licences address and most researchers never check.
The honest review
Strengths. The extraction table is the only feature in the AI-research category that replaces a genuinely expensive human task rather than a cheap one. Reading twenty abstracts is an afternoon; building a comparable table from twenty papers is two days, and Elicit gets you a checkable draft of it in minutes. Retrieval-before-generation also means it does not invent references, which puts it in a different trust class from general assistants.
Limitations. The extractions are machine-made and wrong often enough to require checking — most reliably on explicit, stated fields like sample size, least reliably on anything requiring judgement, such as limitations or the actual nature of a control condition. Coverage follows its underlying corpus, which is strong in biomedicine and the sciences and thin in the humanities and in non-English literature. The credit model can run out mid-review, which is disruptive at exactly the wrong moment. And the interface’s confidence outstrips its reliability: a clean table of extracted values invites you to treat it as data, which it is not until you have checked it.
Verdict. Adopt it for scoping reviews, thesis introductions, and the early stages of any structured review. Skip it as your sole method for a registered systematic review, and skip it in fields where its corpus is thin — check the first page of results for your own topic and you will know within a minute. The condition that flips the answer is whether you will verify: with checking, it is the strongest tool in this category; without, it is an efficient way to publish someone else’s extraction errors.
Common questions
Is Elicit free?
There is a free tier with a monthly credit allowance covering a limited number of searches and extractions, with paid plans for heavier use and institutional options. Credits are consumed by extraction, so light exploratory use goes further than a full review.
Does Elicit hallucinate citations?
Not in the way general assistants do — it retrieves real papers from an academic index before summarising, so the references exist. What it can get wrong is the extraction: the values it pulls from each paper. Verify the fields, not the existence of the paper.
Can I use Elicit for a systematic review?
As an accelerator inside a proper protocol, yes — many teams use it for scoping and for a first extraction pass. As the reported search method, no: it does not produce a reproducible query, and a methods reviewer will ask for one. Run and report a database search as well.