Elicit

Research Methodology Terms Used Here

Five terms that make the difference between understanding this tool and misreading its claims.

Systematic review

Answering a research question by finding and appraising all relevant evidence against a protocol declared in advance. The highest standard of evidence synthesis.

Recall

The share of genuinely relevant papers a search actually found. In systematic review this is the metric that matters, because a miss can change the conclusion.

Screening

Deciding which retrieved records are relevant, first from titles and abstracts and then from full text. Conventionally done independently by two reviewers.

Data extraction

Pulling specified variables out of each included study into a structured table — sample sizes, outcomes, effect sizes and study characteristics.

Publication bias

The tendency for studies with positive results to be published more often, which distorts any synthesis based only on published literature.

What Elicit Is Built To Do

• Elicit is an AI research assistant that searches, screens, summarises and extracts data from academic papers, designed around systematic review methodology rather than around general question answering.
• Coverage is more than 125 million papers, with the vendor stating it can analyse up to 1,000 relevant papers and process up to 20,000 data points in one piece of work.
• Every generated claim carries a sentence-level citation to the source material — the single most important design decision in a research tool.
• It publishes benchmark numbers against 994 Cochrane reviews: 95% search recall, 97% abstract screening, 99% full-text screening and 96% extraction, with data extraction accuracy stated at 99.4%.
• The vendor states more than 5 million researchers use it, across pharmaceuticals, academia, medical devices, policy and industry.

Coverage, Accuracy And Scale

Why Systematic Review Is The Right Problem To Attack

Elicit is not built to answer questions. It is built around a specific and enormously expensive research method, and understanding that method explains the whole product.

A systematic review is the process of answering a research question by finding and assessing all the relevant evidence, according to a protocol declared in advance. It underpins clinical guidelines, drug approvals and policy decisions, and it is the highest form of evidence synthesis there is.

It is also brutal to conduct. A team designs a search, runs it across databases, and retrieves several thousand records. Two reviewers independently screen every title and abstract. The survivors are read in full, again independently. Data is extracted from each included study into a structured table. Disagreements are resolved by a third reviewer.

That process routinely takes twelve to twenty-four months and consumes hundreds of hours of skilled time — most of it spent reading abstracts of papers that turn out to be irrelevant.

Faster tools like Consensus answer a question; this one builds a defensible search. It is exactly the shape of work that should be automated: high-volume, rule-governed, exhausting, and where the human contribution to each individual decision is genuinely small even though the aggregate matters enormously.

That Elicit organises itself around this rather than around a chat interface is the strongest signal about who built it and for whom.

Screening four thousand abstracts by hand is not research. It is the tax you pay before the research, and it is the right thing to automate.

The Cochrane Benchmark Is An Unusually Serious Claim

Most AI tools publish accuracy figures that cannot be checked. This one is different in a way worth explaining.

Cochrane reviews are the gold standard of evidence synthesis in medicine. They are conducted to a published protocol, by trained reviewers, with documented inclusion and exclusion decisions for every record considered.

That makes them a genuine benchmark rather than a marketing exercise. If you can obtain 994 completed Cochrane reviews, you have 994 cases where humans already decided, transparently, which papers were relevant and what data they contained. You can run a system against the same starting point and compare.

The stated results — 95% search recall, 97% abstract screening, 99% full-text screening, 96% extraction — describe agreement with expert human reviewers on tasks where the correct answer was independently established.

The figure that matters most is search recall at 95%, because in systematic review a miss is the defining failure. Missing 5% of relevant literature is not a small error — a review that omits a relevant trial can reach the wrong conclusion, and that conclusion may inform treatment.

So the honest reading is that 95% is genuinely impressive for automation and is not sufficient on its own for a review that will be published or submitted. Which is precisely why it is used to accelerate the process rather than replace the second reviewer.

Sentence-Level Citation Is The Feature That Makes It Usable

If you take one thing from this listing, take this: the citation architecture is more important than the corpus size.

General assistants like Perplexity cite the open web; academic work needs more. Language models fabricate references. This is not an occasional bug but a well-documented characteristic — plausible author names, plausible journals, plausible years, describing papers that do not exist. It is the single reason general chat models are dangerous for literature work, and it has already produced retractions and court sanctions in other fields.

A researcher who cannot instantly verify a claim has to check every reference by hand, which eliminates the time saving that motivated using the tool in the first place.

Grounding every generated sentence in a specific passage of a specific retrieved paper changes that entirely. The claim is traceable, and the check takes a click rather than a search.

It also changes the failure mode in a useful direction. A grounded system that cannot find support tends to say less rather than inventing more — which is the correct behaviour for research and the opposite of what a general assistant does.

The corollary is that you must still read the cited papers before relying on them. A citation proves the source exists and was retrieved; it does not prove the summary characterised its findings correctly, and characterisation is where the remaining errors live.

1. Run a review question you have already completed manually and compare what it retrieves.
2. Look specifically at what it missed — recall is the metric that matters, not what it found.
3. Click through a dozen citations and confirm each paper says what the summary claims.
4. Test a question where the literature is contested, and see whether it reflects the disagreement.
5. Keep it as a first-pass screener alongside a human reviewer before trusting it unsupervised.

How Access Is Sold

This site does not publish plan prices. Elicit's structure is tiered, and the shape is worth knowing because it differs by who you are.

There is a free entry point with paid tiers above it, and separate institutional arrangements for universities, pharmaceutical companies and government bodies.

The meter in research tools of this kind is usually work volume rather than seats — how many papers a run can analyse, how many extractions or reports you can perform, and whether the heavier systematic review workflows are available at your level. Screening 1,000 papers is a materially different unit of consumption from asking a question.

For individual academics the practical question is whether your institution already has an arrangement, since library-negotiated access is common for research tools and much cheaper than an individual subscription.

For commercial research teams the comparison worth running is against the fully loaded cost of the analyst hours currently spent screening. A systematic review consuming several hundred hours of skilled time makes the arithmetic unusually clear.

One thing to confirm regardless of tier: what happens to your uploaded documents and your queries, particularly if you are working on unpublished research or commercially sensitive questions.

What It Cannot Do, Stated Plainly

Being specific here matters more than usual, because the accuracy figures invite over-trust.

It cannot assess study quality the way a methodologist does. Risk of bias assessment — was randomisation adequate, was blinding maintained, was the outcome measure appropriate — requires domain judgement about how a study was actually conducted. Extraction finds what a paper reports; it does not evaluate whether the paper should be believed.

It sees published literature only. Publication bias is one of the most serious distortions in evidence synthesis: studies with negative results are systematically less likely to be published. A perfect search of published papers still inherits that bias, and grey literature and unpublished trial data sit largely outside any corpus.

Cross-checking against Semantic Scholar is a cheap sanity test. Coverage is not universal. 125 million papers is enormous and it is not everything — non-English literature, older material, some regional journals and paywalled content vary in how well they are represented.

The synthesis is still yours. Deciding what a body of evidence means, where studies conflict and why, and what should follow from it, is the actual intellectual work. Elicit gets you to the point where that work can begin, much faster.

For anything that will be published, submitted to a regulator, or used to inform clinical practice, treat this as a powerful assistant to a properly conducted review and not as the review itself.

• Does not perform risk-of-bias or methodological quality assessment
• Inherits publication bias, since unpublished negative results are not there to find
• Coverage varies for non-English, older and paywalled literature
• A citation proves a source exists, not that it was characterised correctly
• Interpretation and synthesis remain the researcher's work entirely