Elicit (AI Research Assistant)

Elicit is a specific, named AI research-assistant platform — not a generic term for “AI that helps with research.” It is built by the public benefit corporation Elicit (which spun out of Ought, a nonprofit machine-learning research lab, in 2023) and searches an indexed corpus of academic literature (elicit.com states over 138 million papers, as of mid-2026) to perform structured evidence-extraction tasks: summarizing individual papers, extracting and tabulating specific data points (methods, sample sizes, populations, outcomes) across many papers at once with sentence-level citations back to source text, and supporting abstract- and full-text screening workflows modeled on systematic-review methodology, including screening aligned to PRISMA 2020 reporting conventions. It sits in the “literature summarization and evidence-synthesis” category of AI-powered research assistant tools — distinct from citation-graph discovery tools like Connected Papers or ResearchRabbit, and from general-purpose AI writing assistants.

What Elicit actually does

As of mid-2026, Elicit’s product surface (per elicit.com) includes several distinct features that are often conflated under a single “AI research assistant” label:

• Research Reports — a structured research brief across a set of retrieved papers, generated by a process the vendor describes as “inspired by systematic reviews,” with the underlying papers and coverage adjustable by the user.
• Systematic review workflows — dedicated screening (title/abstract and full-text) and data-extraction tooling built around the mechanics of a formal systematic review, rather than a one-off summary.
• Tables/extraction — a spreadsheet-style interface where a user defines columns (a variable to extract, e.g. sample size or primary outcome) against a set of source papers; each cell links back to the specific sentence in the source paper it was extracted from.
• Library and alerts — project-level storage for collected sources plus recurring alerts for newly published matching literature.

Access is tiered: a free plan offers unlimited search and paper summaries with limited usage of the systematic-review and report features; paid individual tiers (named “Pro” and “Scale” on elicit.com as of mid-2026) raise the screening/extraction volume limits and add features like figure extraction and API access; an Enterprise tier adds institution-level security controls and higher screening/extraction ceilings. Exact pricing and limits change over time — verify current figures directly against elicit.com/pricing rather than treating any published number as fixed.

The verification gap: what Elicit reports about itself vs. what independent studies found

This is the single most important thing for a research administrator or systematic-review team to understand before relying on Elicit’s output, and it is a genuinely instructive case study in why AI-assisted literature review requires independent verification rather than trusting a vendor’s own accuracy claims:

• Elicit’s self-reported figures (per elicit.com, fetched mid-2026): validation against 994 Cochrane systematic reviews reporting approximately 95% search recall, 97% abstract-screening accuracy, 99% full-text-screening accuracy, and 96% extraction accuracy, plus a separately cited 99.4% data-extraction accuracy figure from vendor case studies.
• An independent peer-reviewed comparison (Lau O, Golder S. “Comparison of Elicit AI and Traditional Literature Searching in Evidence Syntheses Using Four Case Studies,” Cochrane Evidence Synthesis and Methods, 2025, 3(6):e70050) tested Elicit’s search function against the actual search strategies of four completed systematic reviews and found average sensitivity of roughly 39–40% (range approximately 25.5–69.2%), compared with roughly 94.5% sensitivity for the reviews’ original, traditionally conducted searches — while Elicit’s precision (roughly 42%) was notably higher than the traditional searches’ precision (roughly 7.6%). The authors’ conclusion: Elicit did not search with high enough sensitivity to replace a traditional systematic-review search, but its high precision makes it a plausible supplementary or preliminary search tool, not a primary one.
• A separate independent feasibility study (Lagisz M, Mizuno A, Morrison K, Pollo P, Ricolfi L, Yang Y, Nakagawa S, “Using Elicit AI research assistant for data extraction in systematic reviews: a feasibility study across environmental and life sciences,” Research Synthesis Methods, 2026) tested Elicit’s data-extraction feature on 70 variables across seven systematic reviews. Roughly 78% of variables met an 87% accuracy threshold during initial prompt development, but that fell to about 69% when applied to new, previously unseen articles. Re-running the same extraction under different user accounts produced the same extracted value about 90% of the time, but the supporting quote matched only about 46% of the time and the stated reasoning matched only about 30% of the time — meaning the tool can converge on the same answer for different, unstated reasons. The study also found Elicit cannot extract data presented only in figures or tables, and recommends using it as a secondary reviewer or “sanity check” alongside human extraction, not as an autonomous replacement.

None of this means Elicit is unreliable in an absolute sense — both independent studies found genuine value in specific, bounded uses (supplementary search, a second-pass sanity check on extraction). It means the vendor’s own validation figures, run on Elicit’s own chosen review set, measured something different from what happened when independent researchers applied the tool to their own review questions and extraction tasks under realistic conditions. That gap is the reason every output needs the same treatment described in AI in literature search: check the tool, version, and query strategy used, and verify extracted values and cited sentences against the original source rather than the tool’s summary of it.

Hallucination and citation risk

Like every large-language-model-based tool, Elicit is subject to hallucination — generating a plausible-sounding but incorrect summary, misattributing a finding to the wrong paper, or (less commonly, given Elicit’s sentence-level citation design) citing a source that does not actually support the stated claim. Elicit’s per-cell sentence-level citations are a real mitigation — they let a user check the exact source sentence behind an extracted value rather than trusting an unsourced summary — but a citation being present is not the same as the citation being correct; the independent studies above found real gaps between the extracted value and its stated supporting evidence even when a citation was shown.

Disclosure and authorship

Elicit cannot be listed as an author or co-author of a manuscript or review under any major publisher or editorial body’s policy — see AI as author and AI co-authorship rejection (ICMJE 2023). Its use in a published systematic review or manuscript is a disclosable methodological step under current publisher policy once it substantively shapes the search strategy, screening decisions, or data extraction — see CASRAI’s publisher policy landscape for generative AI in manuscripts and, for CRediT-statement mechanics specifically, how to disclose AI assistance in a CRediT statement, role by role. Confidential or unpublished material (an in-progress manuscript, an unregistered review protocol) should not be pasted into Elicit, or any similar public AI tool, without first checking its data-retention and training-use terms.

Examples

• A systematic-review team uses Elicit’s screening workflow to get a first-pass include/exclude recommendation on 200 candidate abstracts against the review’s pre-registered eligibility criteria, then has two human reviewers independently confirm every recommendation before it counts toward the review’s PRISMA flow diagram — using Elicit as a first-pass filter, with full human adjudication, consistent with the “secondary reviewer” role the independent feasibility studies above recommend.
• A researcher builds an Elicit extraction table with a defined research question and a list of 40 papers, asking it to extract sample size, population, and primary outcome into columns, then spot-checks a sample of cells against the underlying PDF before using any value in a manuscript.

Counter-example

A researcher pasting a single paper’s abstract into a general-purpose chatbot and asking for a one-paragraph summary is not “using Elicit” — that workflow is possible in any general large-language-model chat interface and lacks Elicit’s indexed academic corpus, sentence-level source citation, and the screening/extraction workflow purpose-built around systematic-review methodology that defines the product.