AI Research Assistants 2026 — Elicit, Consensus, Scite

An AI research assistant is a search tool that reads papers for you and answers in sentences instead of a list of links. Elicit extracts data into tables, Consensus summarizes what studies conclude, Scite shows how later papers treated a claim, and NotebookLM answers questions about documents you upload. They save real effort at the discovery stage. They also fail in specific, measurable ways, so this comparison covers what each does, what it costs in October 2026, and how to check its output.

The short answer: which tool for which job

The table matches tasks to tools using each vendor's own description of what the product does. It is not a ranking, and nobody has run a neutral benchmark across all of them.

If you need to	Look first at	Why
Screen many papers and extract study details into a table	Elicit	Built for literature reviews; Pro screens up to 5,000 papers
See what the literature broadly says on a yes/no question	Consensus	Offers a "Consensus Meter" showing the level of agreement
Check whether later papers support or contradict a claim	Scite	Smart Citations classify how citing papers treat earlier work
Ask questions about papers you already have	NotebookLM	Answers are grounded in your uploaded sources, with citations
Browse a free index of academic papers	Semantic Scholar	Free, run by the Allen Institute for AI
Map the neighborhood around one key paper	Connected Papers	Builds a graph of related work from Semantic Scholar data
Quick general questions with web sources	Perplexity or another AI search engine	Broader sources, variable quality

A sensible default for most students is free tools first: Semantic Scholar for search, NotebookLM for the papers you have chosen, and a paid plan only once a specific task justifies it. Our companion pieces on AI tools for students and AI search engines cover the surrounding ground.

What these tools actually do: retrieval first, writing second

Most research assistants combine two steps. First, a search system finds papers that match your question. Second, a language model reads the retrieved text and writes a summary or answer. This pattern is called retrieval-augmented generation, covered in depth in What Is RAG?.

Think of a librarian and a novelist. A librarian fetches the books that match your question and tells you what is in them. A novelist produces a convincing story whether or not any book supports it. A research assistant should behave like the librarian, and the retrieval step is what keeps the model tied to real documents. When the second step drifts from the retrieved text, the tool behaves like the novelist: a fluent paragraph that no paper says.

That structure explains the best and worst behaviors of these products. Tools that cite a specific sentence in a specific paper can be checked in seconds. Tools that cite a paper in general terms, or that answer from the model's memory, cannot. Our explainers on why AI hallucinates facts and why AI search tools cite sources differently go into the mechanics.

Elicit

Elicit is aimed at people doing literature reviews. Its homepage says Paper Search can surface up to 1,000 relevant papers at once, that it can analyze up to 20,000 data points at once in interactive tables, and that it automates screening and data extraction for systematic reviews. It also produces research reports with sentence-level citations, a library for storing sources, and "Routines" that track new research (Elicit).

Corpus. Elicit's pricing page says it searches "more than 138 million papers," with unlimited search on the free plan. Older guides quote 200 million; the vendor's current figure is the one to use.

Pricing. According to its pricing page in October 2026, Basic is free, with limited use of the Research Agent and Reports and unlimited paper search, summaries, and chat with full-text papers. Pro costs $49 per month ($588 per year) and screens up to 5,000 papers, extracts 20 table columns at a time, and includes API access. Scale costs $169 per month ($2,028 per year) with five times the usage limits and real-time team collaboration. Enterprise is custom, with screening of up to 40,000 papers (Elicit pricing). The $10 per month plan that older articles list no longer appears.

Vendor claims. Elicit says its case studies show 99.4% data-extraction accuracy and that researchers save up to 80% of their time on systematic reviews. Both are the company's own figures, so treat them as vendor claims until an independent evaluation reproduces them. The reliability section below explains how to run your own spot-check.

Where it fits. Elicit suits empirical fields where papers report samples, methods, and outcomes in a regular structure. It is a weaker fit for theoretical or humanities work, where there are no sample sizes to extract.

Consensus

Consensus presents itself as an AI search engine for peer-reviewed literature. Its site lists Pro Search and Deep Search modes that return AI-synthesized answers; the Consensus Meter, which shows the level of scientific agreement on yes/no questions; Ask Paper, for chatting with a specific paper; Study Snapshots with key findings and methodology; a Citation Graph; and Collections. Citations can be exported in RIS, BibTeX, and CSV formats, and the page describes a free tier. It says it searches over 400 million scientific papers (Consensus).

A note on corpus counts. Elicit says 138 million, Semantic Scholar says over 200 million, and Consensus says over 400 million. These numbers are not comparable. They depend on what each vendor counts as a paper, whether it holds full text or only abstracts, and how it removes duplicates. Do not pick a tool because its number is larger.

Pricing. Consensus advertises a free tier. Paid plan prices change often, so check the vendor.

The yes/no trap. A meter that summarizes "yes," "no," and "possibly" is convenient, and it invites a specific mistake: treating the count of studies as the weight of evidence. A tally mixes small and large studies, strong and weak designs, and association with causation. The articles on correlation versus causation and why randomized trials are powerful explain why a vote count of papers can mislead. Use the meter to find the studies, then read them.

Scite

Scite's distinctive feature is the Smart Citation. Instead of only counting how often a paper has been cited, it shows the sentence in each citing paper and classifies how the citing paper treats the claim. Scite's site says this shows whether studies "support or contradict a claim," and it describes the categories as supporting, challenging, and discussing. Its homepage cites 1.6 billion or more citations indexed, over 300 million articles, preprints, books, patents, and datasets in search, and more than 2 million users (Scite).

Pricing. When billed yearly, Basic is $12 per month, with full Assistant and Search access, Smart Citation reports, and collections up to 1,000 papers. Pro is $40 per month and adds API access, larger collections, and patent, clinical trial, and grant datasets. Both include a 7-day free trial (Scite pricing).

A claim worth testing. Scite says its Assistant gives answers "grounded in real papers" and that references are never generated or hallucinated. That is the vendor's statement, and its page offers no independent verification. The practical response is the verification routine below: click the cited sentence and read the paper around it.

Where it fits. Scite is most useful late in a project, when you hold a claim and need to know whether it survived scrutiny. A citation count of 400 says little. A breakdown showing many later papers challenging the finding says a lot.

NotebookLM

NotebookLM is Google's tool for working with documents you supply. Its site describes answers grounded in your sources, with citations, and Audio Overviews. Google's help page lists limits that vary by plan: the Standard (free) plan allows 100 notebooks per user and 50 sources per notebook, Plus allows 200 notebooks and 100 sources, and Pro allows 500 notebooks and 300 sources, with higher tiers above that, and Google notes the limits can change (Google NotebookLM help).

Plan	Notebooks per user	Sources per notebook
Standard (free)	100	50
Plus	200	100
Pro	500	300

Strengths. Because the tool only reads what you upload, it cannot invent a paper that is not in the notebook, and each answer can point back to a passage. It works well for one project's reading list or a stack of course readings. Our guide How to Use NotebookLM walks through the workflow.

Limits. It is not a search engine. It will not find the paper you did not know existed, and it inherits any weakness of the source file. Scanned PDFs, tables, and equations can be misread, a problem discussed in why AI PDF tools struggle with complex layouts. For a broader look at document chat tools, see Best AI PDF Tools 2026.

Semantic Scholar and Connected Papers

Two lower-cost tools deserve a place in the stack.

Semantic Scholar is a free research tool built by Ai2, the Allen Institute for AI, a non-profit founded in 2014 by Microsoft co-founder Paul Allen. It launched in 2015, indexes over 200 million academic papers, and provides AI-driven search and discovery along with a free API and datasets (Semantic Scholar). For many literature searches it is the best free starting point.

Connected Papers draws a graph of papers related to one you already know. Its pricing page lists Free, Academic, and Business tiers and names Semantic Scholar as its data source (Connected Papers). Starting from one important paper and seeing its neighbors is a good way to find work that keyword searches miss.

Perplexity and general-purpose assistants

General assistants and AI search engines can answer research-flavored questions, but their sources are broader than peer-reviewed journals, and quality varies. We compare them in Perplexity vs ChatGPT Search vs Google AI Overviews and Best ChatGPT Alternatives 2026. For long, multi-step investigations that browse many pages, see our explainer on deep research AI tools.

The next section shows why the distinction between a research-specific tool and a general chatbot matters.

How reliable are the citations?

Reliability is the question that decides whether any of these tools belongs in your workflow. Two sets of measurements are public, both about general-purpose chatbots, not the specialist tools above.

Evidence from 2023. Walters and Wilder asked ChatGPT-3.5 and ChatGPT-4 to write short literature reviews on 42 topics, producing 84 papers and 636 citations, then checked each citation against databases and websites. In their results, 55% of GPT-3.5 citations and 18% of GPT-4 citations were fabricated, and among citations to real papers, 43% and 24% respectively contained substantive errors (Scientific Reports, 2023).

Evidence from 2025. A study posted to arXiv tested eight free-tier chatbots (ChatGPT, Claude, Copilot, DeepSeek, Gemini, Grok, Le Chat, and Perplexity) on 400 academic references across five fields. Overall, 26.5% of references were fully correct, 33.8% were partially correct, and 39.8% were erroneous or fabricated. Grok and DeepSeek were the only chatbots that produced no false references, while Copilot, Perplexity, and Claude had the highest hallucination rates. Journal articles had a significantly higher fabrication rate than book references (arXiv 2505.18059).

What these numbers do and do not show. Both studies tested general chatbots asked to produce references. They did not test Elicit, Consensus, Scite, or NotebookLM, which retrieve real documents first. The vendors argue that retrieval removes the fabrication problem, and Scite says its references are never generated. That may be largely true, but nobody should assume it. A tool can retrieve a real paper and still misstate what it found, cite the wrong section, or summarize an abstract as if it were a result. The risk moves from "this paper does not exist" to "this paper does not say that."

A verification routine that takes five minutes

For any claim that will appear in your own work, the following checks catch most problems.

1. Open the cited paper. Confirm that it exists by clicking through to the publisher or an index such as Semantic Scholar, not by trusting the tool's own page.
2. Find the sentence. Locate the passage the summary relies on. If the tool names a sentence, read it in context. If it does not, search the paper for the key term.
3. Check the study type. Was it a randomized trial, an observational study, a review, or a case report? The same sentence means different things in each.
4. Check the numbers. Compare any sample size, effect size, or percentage the tool reports with the paper's tables.
5. Look for contradiction. Use a citation-context tool such as Scite, or a search for later critiques, to see whether the claim has been challenged.
6. Record the primary source. Cite the paper, not the tool.

Spot-checking a sample is a reasonable standard for a vendor's data-extraction claims too. If you extract 20 fields from 40 papers, pull five papers at random and verify every field by hand. If errors cluster, widen the check.

Literature reviews and systematic reviews: what the guidance says

A narrative literature review and a systematic review are different jobs. A systematic review follows a pre-registered protocol and fixed rules for which studies to include, which makes it much more sensitive to a tool's mistakes.

In November 2025 Cochrane, the Campbell Collaboration, JBI, and the Collaboration for Environmental Evidence published a joint position statement on AI in evidence synthesis, endorsing the RAISE recommendations. Its main rules are worth knowing even if you never write a systematic review (position statement in JBI Evidence Synthesis, via PMC):

• Synthesists carry final responsibility for their work, including the decision to use AI.
• AI may be used if authors can show it does not compromise methodological rigor, and it should be used with human oversight.
• Any AI that makes or suggests judgments, such as eligibility decisions, risk-of-bias appraisals, data extraction, or certainty ratings, must be fully and transparently reported.
• Reports should state the tool's name, version, and date, its purpose, its justification and validation, and its limitations and biases.
• Tool developers are asked to publish how their tools work and to make validation evaluations public.

The statement treats large language models and generative AI under the same rules, and it flags fabricated outputs and opaque limitations of commercial tools as specific concerns. For a student or researcher, the takeaway is that a tool's speed does not change who is accountable. If Elicit screens 5,000 papers and excludes 4,800 of them, the excluded papers still need a documented, checkable rationale.

A research workflow that uses the tools honestly

Popular claims that these tools cut a topic review from dozens of hours to a handful, or shrink a systematic review from months to weeks, have no independent study behind them. What can be said is that the tools change where time goes. The typical stages, with the role of each tool, look like this:

Stage	What you do	Where AI helps	What stays human
Frame the question	Write a precise, answerable question	Brainstorming search terms	Deciding what matters
Find papers	Search several indexes	Elicit, Consensus, Semantic Scholar, Connected Papers	Choosing databases; checking coverage gaps
Screen	Apply inclusion and exclusion criteria	Elicit's screening; Consensus summaries	Final inclusion decisions; documenting reasons
Extract data	Pull sample sizes, methods, outcomes	Elicit tables	Verifying values against the papers
Read key papers	Understand methods and limits	NotebookLM questions on your sources	Critical judgment of design and bias
Check claims	See how later work treated findings	Scite Smart Citations	Weighing contradictory evidence
Write	Synthesize and cite	Drafting help, within journal policy	Argument, interpretation, and the final text
Organize	Keep references and notes	Reference managers	Consistent notes; see note-taking systems for researchers

A good habit for the "read key papers" stage is to read the paper yourself first and then use the tool to test your understanding, not the other way around. Our guide on how to read a research paper without getting lost covers the manual method that the AI tools should supplement.

Four concrete uses

A quick fact-check. You see a claim that a supplement reduces symptoms. Ask Consensus or a similar tool for what studies conclude, then open the two or three most relevant studies and look at their designs. A statement supported by randomized trials weighs more than one supported only by observational data.

Decoding a hard paper. Upload the PDF to NotebookLM and ask for the main findings, the methodology, and the limitations, then check the answers against the paper's abstract and discussion section. Ask it to point to the passage behind each answer.

Finding what you missed. Start from one landmark paper in Connected Papers and see which neighbors you have not read. Then look up each in Semantic Scholar to confirm it exists and check its citation context.

Testing a citation. You find a sentence in a review reading "X causes Y (Smith 2019)." Use Scite to see how Smith 2019 has been cited, then read the original. Reviews sometimes cite a paper for a claim it only partly supports.

What these tools cannot do

• Replace deep reading. A summary compresses a paper and drops detail, including the detail that matters.
• Supply judgment. Deciding whether a study's design supports its conclusion is the researcher's job.
• See what is not in their index. Paywalled full texts, unpublished work, dissertations, and non-English literature may be missing or only partly covered, and index coverage differs by vendor.
• Guarantee currency. New papers take time to be indexed, so supplement with a direct database search for fast-moving topics.
• Protect confidential data. Check a tool's data handling before uploading unpublished manuscripts or sensitive material. Read each vendor's terms for your plan.

What it costs

Prices below come from each vendor's page in October 2026. Where no price is given, check the vendor.

Tool	Plan	Price	Notes
Elicit	Basic	Free	Unlimited paper search; limited agent and reports
Elicit	Pro	$49 per month ($588 per year)	Screens up to 5,000 papers
Elicit	Scale	$169 per month ($2,028 per year)	5x usage limits; collaboration
Scite	Basic	$12 per month, billed yearly	7-day free trial
Scite	Pro	$40 per month, billed yearly	API access; larger collections
NotebookLM	Standard	Free	50 sources per notebook
Semantic Scholar	Free	$0	Free API and datasets
Consensus	Free tier exists	Check vendor	Paid plans vary
Connected Papers	Free, Academic, Business	Check vendor	Tiers listed on its pricing page

A student who adds a paid Elicit Pro plan ($49) and a Scite Basic plan ($12) is already spending $61 per month before any general chatbot subscription. That is reasonable for a funded project and hard to justify for casual reading. Many universities and libraries license research tools, so check with yours before paying personally.

Ethics, authorship, and disclosure

Authorship. The International Committee of Medical Journal Editors (ICMJE) says AI tools should not be listed as authors, because they "cannot be responsible for the accuracy, integrity, and originality of the work." Authors must state at submission whether they used AI tools and explain how, and nondisclosure may lead to corrective action and in some cases be treated as misconduct (ICMJE).

Responsibility. The same guidance says humans remain accountable for any submitted material that involved AI. Authors should check and edit AI output, since it may be wrong, incomplete, or biased, and must be able to confirm the work contains no plagiarism.

Citation. ICMJE says citing AI-generated content as a primary source is unacceptable. Cite the paper that contains the finding, not the tool that led you to it. Journal and institutional policies vary beyond that baseline, so read the policy for your target venue.

Plagiarism. Pasting a tool's summary into your draft without attribution is a problem even when the summary is accurate. Read, understand, write in your own words, and cite the original.

Confidentiality. Unpublished data, peer-review materials, and participant information need the strictest handling. Use an institutional account with a clear privacy agreement, or keep that material out of cloud AI tools.

FAQ

Q: Can AI write my paper for me? A: It can assist with parts of the process, but accountability stays with you. ICMJE guidance says AI tools cannot be authors, that AI use must be disclosed at submission, and that authors must verify the output. Journals set their own policies, so check the one you are targeting.

Q: Are AI research tools accurate enough to trust? A: Not without checking. Tests of general chatbots found fabricated references at high rates: 55% for GPT-3.5 and 18% for GPT-4 in a 2023 study, and 39.8% erroneous or fabricated across eight chatbots in 2025. Retrieval-based tools such as Elicit, Consensus, and Scite are designed to reduce this, but their vendors' accuracy claims are not independently verified in the sources reviewed here.

Q: What is the best free option for serious research? A: A combination. Semantic Scholar is free and indexes over 200 million papers, Elicit's Basic plan includes unlimited paper search and summaries, and NotebookLM's free plan allows 50 sources per notebook. Use Google Scholar or your library's databases alongside them.

Q: Is Elicit or Consensus better? A: They solve different problems. Elicit is built around extracting structured data from many papers for literature reviews. Consensus is built around summarizing what studies conclude on a question. Many researchers try both on a question they already know the answer to and compare.

Q: Can I use these tools for a systematic review? A: Possibly, with conditions. The joint Cochrane, Campbell, JBI, and CEE position statement of November 2025 says AI may be used with human oversight, that its use must be justified and reported transparently, and that tool name, version, date, validation, and limitations should be stated. Follow the protocol of your target journal.

Q: How should I cite AI-assisted work? A: Cite the original sources you read and verified, not the tool. Disclose AI use as your journal or institution requires. ICMJE says citing AI-generated content as a primary source is unacceptable.

Q: Do these tools replace Google Scholar? A: No. Treat them as additions. Their indexes differ in coverage and age, and each vendor describes its corpus differently, so searching more than one index reduces the chance of missing a key paper.

Related Articles

• How to Use AI for Studying — A Practical Guide (2026)
• Why Headlines Overstate Breakthrough Science

The takeaway

AI research assistants are best understood as fast first readers. They find candidates, line up details, and point to passages, which moves hours of drudgery into minutes. They do not move accountability. A fabricated or misread citation in your paper is yours, whatever tool produced it, and the measured error rates for general chatbots show why a five-minute check is cheap insurance.

The sensible stack is small and mostly free: a search index, a tool for reading your own sources, and a paid specialist only when a task demands it. Learn the verification routine before you learn any tool's advanced features, since the routine still works after the tools change. Prices, corpus sizes, and plan names in this field move often, so treat every figure here as dated October 22, 2026.

Last verified: October 22, 2026