Elicit

What is Elicit?

Elicit is an AI research assistant built around a single premise: that the bottleneck in modern science is not generating ideas but reading the evidence that already exists. Where a general-purpose chatbot will answer a research question from whatever it absorbed during training, Elicit does something narrower and more auditable — it searches a defined corpus of academic literature, pulls the specific papers that bear on your question, and lays their contents out as a table you can inspect column by column. The product is aimed at people who have to defend their sources: systematic reviewers, evidence synthesis teams, regulatory and HTA analysts, pharmaceutical and medical device researchers, policy staff, and the graduate students who do the unglamorous screening work underneath all of it.

The company behind it has an unusual lineage. Elicit began inside Ought, a non-profit research organization, and became an independent public benefit corporation in September 2023 alongside a $9 million seed round led by Fifty Years, with angel participation from figures including Google's Jeff Dean and GitHub co-founder Tom Preston-Werner. A $22 million Series A followed in February 2025 at a $100 million valuation, co-led by Spark Capital and Footwork. That origin matters to how the product behaves: the founding thesis at Ought was about supervising the process of machine reasoning rather than just checking its outputs, and Elicit's insistence on showing the sentence behind every claim is that thesis rendered as a user interface.

Positioned against neighbours in the same category, Elicit sits at the structured-data end. Tools like Consensus optimise for quickly reading the state of evidence on a yes-or-no question; Semantic Scholar is a discovery layer over a citation graph. Elicit's centre of gravity is extraction — turning a pile of PDFs into rows and columns that can be filtered, sorted, exported to CSV, and dropped into an analysis. That focus is also where the independent evidence about its performance is most concentrated, and where its real limitations live.

Core Features

• Structured data extraction into tables: This is the capability the product is built around, and Elicit describes it plainly as one of its superpowers, including extraction from tables and figures where, as the company notes, most clinical outcomes actually live. You define the columns you want — sample size, intervention, outcome measure, effect direction, funding source — and Elicit populates them across every paper in your set. Both quantitative and qualitative fields are supported, and each cell carries the quote or figure that justifies it.
• Semantic search over roughly 138 million papers: Elicit searches across a corpus assembled from Semantic Scholar, OpenAlex and PubMed, deduplicated and stripped of incomplete listings to roughly 138 million records, with ClinicalTrials.gov available as a separate corpus covering over 545,000 registered trials. Retrieval is not keyword matching. Elicit ranks the corpus by semantic similarity using custom embedding models that also weigh citation count and recency, then takes the top 1,000 papers and re-prioritises them with language models before showing you anywhere from the top four to the top thousand.
• Sentence-level citations: Every AI-generated claim is tied to the exact sentence or figure in the underlying paper, rather than merely to the paper as a whole. The company draws the contrast explicitly, noting that other tools link only to references and leave you to read the whole paper to find where an answer came from — and that sometimes the information is not even in the cited source. This is the single most consequential design decision in the product, because it converts verification from a research task into a glance.
• Systematic review workflow with auditable screening: Elicit runs the full arc from protocol refinement through search, screening, extraction and synthesis. Screening scales to 40,000 papers on the enterprise tier, and every include or exclude decision carries an exclusion reason, per-criterion scores and supporting quotes, which the company frames as PRISMA-auditable. Dual review can pair two humans with AI assistance, or use the AI as the second reviewer outright.
• Research reports across up to 200 papers: The report workflow synthesises evidence across as many as 200 sources into a multi-page literature review, with each claim backed by sentence-level citations and a methods section containing PRISMA flow diagrams, paper summary tables and the search strategy used. Built-in templates can be customised per organisation.
• Library and alerts: Papers you find are stored and organised for reuse across projects, and saved searches can be converted into alerts that surface new publications as they appear — the mechanism behind what Elicit calls living reviews, where a completed review is refreshed rather than rebuilt.
• Bring your own documents and subscriptions: You can upload your own PDFs for custom extraction, and connect institutional journal subscriptions to widen full-text access beyond the open-access subset.

Use Cases

1. Systematic reviews and evidence synthesis: A conventional systematic review by a five-person team is a months-long undertaking dominated by screening and extraction labour. Teams use Elicit to run screening across thousands of records with auto-suggested criteria they can edit, then extract structured fields into a table for analysis. The auditable trail — exclusion reasons, per-criterion scores, supporting quotes — is what makes the output defensible in a methods section rather than merely fast.
2. Scoping and preliminary searches before committing to a protocol: Defining a review protocol before you know what literature exists is a chicken-and-egg problem. Because Elicit's retrieval is semantically driven and unusually precise, it is well suited to rapid reconnaissance: testing whether a question has been studied, how it has been operationalised, and where the disagreements are, before investing in a full search strategy.
3. Pharmaceutical, medical device and HTA evidence work: Regulatory and health technology assessment work involves repeatedly assembling comparable evidence tables across trials — outcomes by trial arm, adverse events, progression-free survival. Elicit's extraction from tables and figures targets exactly this, and the company points to industrial users including a consulting firm that applied it to a policy-facing review.
4. Policy briefs and rapid reviews under deadline: When a decision needs an evidence base in days rather than months, a rapid scoping review over a hundred-odd studies becomes feasible with automated extraction plus human verification of flagged items. One methodologist quoted by the company describes completing exactly this kind of rapid scoping review over more than a hundred studies in a single day.
5. Doctoral literature reviews and comprehensive exams: Graduate students use Elicit to map a field — finding the seminal work, tracking how a construct has been measured across studies, and building the comparison table that underpins a literature review chapter. The structured output is markedly more useful for this than a list of links, though as discussed below it should not be the only search method used.
6. Ongoing surveillance of a research area: Combining saved searches with alerts turns a one-off review into a monitored feed, so that a team responsible for a therapeutic area or a technology domain learns about relevant publications without maintaining manual database alerts.

How to use Elicit

1. Create an account at elicit.com. The free Basic tier is usable immediately and gives unlimited search, summaries and chat with papers, so you can evaluate retrieval quality on your own topic before paying anything.
2. Enter your research question in natural language rather than as a Boolean string. Elicit's ranking is semantic, so a well-specified question ("does online CBT match face-to-face CBT for treating depression in adults?") outperforms a keyword soup. Elicit will often suggest ways to sharpen an over-broad question.
3. Review the returned papers and apply screening criteria in natural language to narrow the set. Read the exclusion reasons Elicit gives — this is where you catch a criterion that is being interpreted differently from how you intended.
4. Define your extraction columns. Start with a handful of concrete, factual fields and inspect the supporting quote in each cell before adding more columns; getting the column definitions right on twenty papers is far cheaper than discovering a misinterpretation across four hundred.
5. Verify the extracted values against sources, prioritising interpretive columns over factual ones. Export the resulting table to CSV or XLSX for analysis, or generate a research report if you need narrative synthesis with a methods section.
6. Save the search as an alert if the topic is ongoing, and keep the papers in your library so the screening work can be reused in the next project.

Tips & Best Practices

• Do not use Elicit as your only search method for a comprehensive review. The independent evidence on this point is unambiguous and is discussed in detail below; supplement with traditional database searches in PubMed, Embase, Web of Science or your field's specialist databases.
• Split factual columns from interpretive ones and treat them differently. Extraction is reliable for things like participant counts and reported measures, and markedly less reliable for anything requiring judgement about what a study means. Verify the interpretive columns manually; spot-check the factual ones.
• Use Elicit as a second reviewer rather than a sole extractor. A published proof-of-concept found that where Elicit and a human reviewer agreed, accuracy was 100% — which makes disagreement, not extraction itself, the useful signal. Extracting in parallel and reconciling only the conflicts is both faster and more accurate than either party alone.
• Check whether Elicit actually has the full text before trusting a full-text-dependent field. Without full-text access it falls back to title and abstract, which will silently produce shallower extractions rather than an error. Installing the browser extension and connecting institutional subscriptions materially changes what the tool can see.
• Record your search as a keyword search when reproducibility is required. Conversational and semantic searches are not directly reproducible, which matters when a methods section has to be replicable by a reader.
• Start narrow, then widen. Because retrieval is precise but not exhaustive, running several differently-framed versions of the same question surfaces papers that a single phrasing misses.
• Read the quote, not just the value. The supporting quote is the entire point of the citation design; a cell whose quote does not obviously support the value is the fastest way to catch an extraction error.

Who is Elicit for?

• Systematic reviewers and evidence synthesis specialists: The primary audience, and the one the product's screening and extraction workflow is explicitly engineered around, including PRISMA-auditable decisions and dual review.
• Pharmaceutical, biotech and medical device researchers: Teams assembling trial evidence tables where outcomes by trial arm and adverse event data need to come out of tables and figures, not just abstracts.
• Health technology assessment and regulatory analysts: Users who must produce defensible evidence bases under structured methodological requirements and time pressure.
• Policy researchers and government analysts: People writing evidence-backed briefs on compressed timelines, where a rapid scoping review is the realistic alternative to no review at all.
• Doctoral and postgraduate researchers: Students building literature review chapters, mapping how constructs have been operationalised, and preparing for comprehensive exams.
• Academic librarians and research support staff: Professionals evaluating and teaching AI research tools, several of whom have published independent institutional assessments of Elicit.
• Corporate R&D and competitive intelligence teams: Industrial users applying evidence synthesis outside academia, which the company has signalled as its direction of expansion.
• Clinicians and clinical researchers: Users who need to locate and compare trial evidence, with ClinicalTrials.gov searchable as a distinct corpus alongside the paper index.

Platforms

• Web application: The primary and fully-featured interface, accessed through a browser at elicit.com with no installation required.
• Browser extension: Used to extend full-text access by leveraging journal subscriptions you already hold, which meaningfully widens what Elicit can read beyond open-access papers.
• REST API: Available from the Pro tier upward, with unlimited access on Enterprise. The documented interface models research agent sessions and returns interactive artifacts — tables, prose, presentations and figures — with tables downloadable as CSV or XLSX and prose as Markdown via short-lived presigned URLs.
• Reference manager integration: Import and export with Zotero, and connectivity with EndNote and Mendeley, so found papers flow into whatever citation workflow you already run.
• No native mobile or desktop applications are advertised on the official site; the product is delivered through the browser (To be verified against the official site for any newer clients).

Pricing & Plans

Elicit offers a genuinely usable free tier. Basic costs nothing and provides unlimited search across the full corpus, unlimited paper summaries, unlimited chat with papers where full text is available, source visibility for answers, and Zotero import — while metering the Research Agent and Research Reports to limited usage. For evaluating whether the retrieval quality suits your field, this is enough to make an informed judgement without spending anything.

Paid tiers are separated by the volume of research work they permit rather than by which buttons are unlocked, which is an unusually honest way to structure a research product. Pro runs $49 per month, or $588 annually with a stated 35% saving, and adds standard usage of the agent, reports and systematic review workflows, screening of up to 5,000 papers, twenty table columns, extraction from 135 data sources, ten personalised alerts, custom extractions from uploaded papers, and API access. Scale runs $169 per month, or $2,028 annually at a stated 39% saving, and provides five times the Pro usage allocation plus extraction from figures, live editing and real-time collaboration, 200 data sources, thirty columns and an admin panel with usage tracking and seat management. Enterprise is custom-priced and adds screening of 40,000 papers, forty columns, unlimited alerts and API access, SSO and SAML, 2FA, user analytics, custom integrations and a dedicated customer success team. Exact entitlements and current prices should be confirmed on the official pricing page.

Alternatives

• Consensus: Also an AI search engine over peer-reviewed literature, but optimised for rapidly reading the balance of evidence on a focused question rather than for building extraction tables; a better fit when you want a fast verdict, a weaker fit when you need structured columns.
• Semantic Scholar: A free academic discovery and citation-graph tool from the Allen Institute for AI — and, notably, one of the corpora Elicit itself draws on. Excellent for finding and following papers, but it does not perform structured extraction or screening.
• Traditional systematic review software (Covidence, Rayyan, DistillerSR): Purpose-built for managing the review process with established methodological rigour and team workflows, but with less AI-driven extraction; often used alongside rather than instead of an AI extraction layer.
• Traditional bibliographic databases (PubMed, Embase, Web of Science, Scopus): Not competitors so much as the necessary complement — they provide the exhaustive, reproducible Boolean searching that the independent evidence shows AI semantic search does not yet replace.
• General-purpose AI assistants with research modes: Convenient for orientation and summarisation, but without a bounded academic corpus or sentence-level source anchoring, which makes them unsuitable where citation accuracy must be verifiable.

Limitations & Considerations

• Search sensitivity is substantially lower than vendor benchmarks suggest, and this is the most important thing to know about the tool. Elicit's own published evaluation, run against 994 open-access Cochrane reviews, reports 96.9% abstract-screening sensitivity and 99.5% paper-level recall at the full-text stage. But Lau and Golder, writing in Cochrane Evidence Synthesis and Methods in 2025, found that Elicit's search sensitivity averaged 39.5% across four case-study reviews, against 94.5% for the original searches. In plain terms: used alone, it missed most of the studies a traditional search found. The same study found Elicit's precision averaged 41.8%, dramatically higher than the 7.55% precision of the original review searches — so what you do get back is far less cluttered with irrelevant results. The honest reading is that this is an excellent instrument for precise, exploratory retrieval and an inadequate one for exhaustive retrieval. The authors concluded that Elicit is currently not sensitive enough to replace traditional searching, while noting the product is continually improving and warrants re-evaluation.
• Extraction accuracy degrades on unfamiliar material and is weakest exactly where judgement is required. A feasibility study across environmental and life sciences concluded that Elicit can complement, but not replace, human data extractors, after finding that variables meeting the accuracy threshold fell from 78% during prompt development to 69% when applied to new, previously unseen studies. A 2025 proof-of-concept in Social Science Computer Review compared Elicit against human reviewers across 43 studies and 602 data points, measuring 81.4% accuracy for Elicit versus 86.7% for humans — a difference that was not statistically significant, which is a genuinely strong result. But the breakdown is the useful part: agreement on participant counts reached 81.4%, but agreement on main results fell to 34.8% and on constructs of interest to 16.3%. The same study observed Elicit sometimes extracting from cited studies rather than the primary source. Factual fields are broadly trustworthy with spot-checks; interpretive fields are not, and must be verified.
• Reproducibility is imperfect at both the search and extraction layers. Re-running identical prompts under different accounts reproduced 90% of the extracted values exactly, but only 46% of the supporting quotes and 30% of the reasoning narratives — meaning the answer is usually stable while the justification for it often is not. Separately, Deakin University's library evaluation notes that conversational searches are not directly reproducible and currently lack the transparency required for systematic review reporting. Where a methods section must be replicable, use keyword search and document it as such.
• Full-text access, not corpus size, is the practical ceiling. Full-text analysis works when a paper is open access, or when you have the Elicit browser extension plus your own journal subscription; otherwise Elicit reads only the title and abstract. In fields where the definitive literature sits behind paywalls, this quietly limits extraction depth without announcing that it has done so.
• Coverage has hard edges that are easy to miss. Books, dissertations and non-academic publications are excluded outright, and Elicit's own documentation warns that grey literature and some clinical trial reports have lower retrieval rates in automated searches, with regional gaps including Chinese-language publications. Users in the humanities and social sciences repeatedly report thinner coverage than in science, technology and medicine. Deakin's evaluation additionally flags a risk of English-language bias and the absence of formal critical appraisal functionality — Elicit will extract what a study reports, but it will not tell you whether the study was any good.
• The hallucination question has not been retired, only constrained. TechCrunch's Kyle Wiggers, covering the seed round in September 2023, questioned whether the underlying hallucination problem had been solved at all, invoking Meta's Galactica — a scientific language model withdrawn after three days for fabricating citations. Sentence-level citation is a serious mitigation because it makes fabrication visible rather than preventing it outright, but the burden of looking still sits with the reader. Evidence-synthesis practitioners repeatedly caution that the tool is not reliable enough for formal meta-research without manual verification at the screening stage.
• Vendor-reported figures should be read with their caveats attached. To its credit, Elicit publishes the caveats alongside the numbers: extraction was measured on open-access studies only, and semantic search was evaluated without the full keyword-plus-semantic workflow, with a small full-text screening sample and extraction questions reconstructed after the fact rather than taken from original protocols. Customer case studies quoted on the site — such as VDI/VDE reporting that Elicit correctly extracted 1,502 of 1,511 data points in a systematic review informing German education policy — are vendor-relayed results from favourable conditions, not independent evaluations.
• The published numbers do not agree with each other. The homepage advertises over 125 million papers while the pricing page and help centre both say 138 million, and user counts appear as two million, five million and 400,000 per month in different places. The 138 million figure is the best-documented, being the only one accompanied by an explanation of how overlapping records across the three source databases are deduplicated. The user figures are plausibly measuring different things — cumulative registrations versus monthly actives — but the site does not say so.

FAQ

Q1. Is Elicit free to use?

Yes, there is a free Basic tier that includes unlimited search across the full corpus, unlimited paper summaries, unlimited chat with papers, and Zotero import, while limiting usage of the Research Agent and Research Reports. Paid tiers begin at $49 per month for Pro; check the official pricing page for current terms.

Q2. How is Elicit different from Consensus or Semantic Scholar?

All three search academic literature, but their outputs differ. Consensus is optimised for quickly reading the balance of evidence on a focused question, Semantic Scholar is a discovery and citation-graph tool, and Elicit's distinguishing capability is extracting structured data from papers into customisable tables — including from tables and figures — with each value tied to a supporting quote.

Q3. Can I rely on Elicit alone for a systematic review search?

No. An independent 2025 study in Cochrane Evidence Synthesis and Methods found Elicit's search sensitivity averaged 39.5% versus 94.5% for the original reviews' searches, concluding it is not currently sensitive enough to replace traditional searching. Use it as a supplement to conventional database searches, where its high precision makes it valuable for preliminary and exploratory work.

Q4. How accurate is Elicit's data extraction?

It depends heavily on what you are extracting. A 2025 study measured 81.4% overall accuracy against 86.7% for human reviewers — statistically indistinguishable — but agreement was 81.4% on participant counts and only 16.3% on constructs of interest. Factual fields are broadly reliable; interpretive fields require manual verification.

Q5. Does Elicit hallucinate citations?

Its core mitigation is sentence-level citation: every AI-generated claim is anchored to the exact sentence or figure in the source, so an unsupported claim is visible rather than hidden. This substantially reduces the risk compared with general chatbots, but it does not eliminate the need to read the supporting quote, and independent researchers still advise manual verification for formal evidence synthesis.

Q6. Can Elicit read paywalled papers?

Partially. It indexes both open-access and subscription papers, but full-text analysis requires either an open-access paper or the Elicit browser extension combined with your own journal subscription. Without full-text access, Elicit works only from the title and abstract.

Q7. What sources does Elicit search?

Roughly 138 million papers assembled from Semantic Scholar, OpenAlex and PubMed after deduplication and removal of incomplete listings, plus over 545,000 trials on ClinicalTrials.gov as a separate corpus. Books, dissertations and non-academic publications are not included.

Q8. Does Elicit have an API?

Yes, API access is included from the Pro tier upward and is unlimited on Enterprise. The documented interface covers research agent sessions and returns artifacts including tables downloadable as CSV or XLSX and prose as Markdown.

Q9. Can teams collaborate in Elicit?

Live editing and real-time collaboration are available from the Scale tier, which also adds an admin panel with usage tracking and seat management. Enterprise adds SSO and SAML, 2FA and user analytics.

Q10. Is Elicit suitable for humanities and social sciences?

It is usable but weaker there. The corpus covers all disciplines in principle, but users consistently report thinner coverage in the humanities and social sciences than in science, technology and medicine, and an independent library evaluation flags a risk of English-language bias.