Best AI for Research: Which Tool for Which Job

The question has no single answer because research is several different jobs. For finding and triaging literature, Elicit and Consensus are the strongest general-purpose starting points and Semantic Scholar is the best free one; for judging whether a claim held up, Scite is the only tool in the group built specifically to classify how later papers cited it. For reasoning, restructuring and drafting, a general assistant such as ChatGPT, Claude or Gemini is well ahead of anything purpose-built for academia. For keeping your references correct, Zotero or Paperpile, neither of which is really an AI product and both of which matter more than any AI tool on this page. If you want one recommendation: run a reference manager, use a discovery tool for orientation only, and never let either class be the last thing that checks a citation.

Zotero, and it is not close. It is free, open source, maintained by the Corporation for Digital Scholarship, and it is the only tool discussed here whose output is authoritative rather than generated. Every other tool on this page produces something you then have to check; a reference manager produces the record you check against. Add an AI layer to it later if you want one — the plugin ecosystem is real but uneven, and plugin compatibility routinely breaks across Zotero major versions.

These are two different questions with two different answers. Peer review: NIH prohibits it outright. Notice NOT-OD-23-149, "The Use of Generative Artificial Intelligence Technologies is Prohibited for the NIH Peer Review Process", states that reviewers are prohibited from using AI tools in analysing and critiquing NIH grant applications and R&D contract proposals, on confidentiality grounds — generative platforms offer no assurance about how uploaded content is stored or reused. NSF, in its Notice to the Research Community of 14 December 2023, likewise prohibits reviewers from uploading any proposal content, review information or related records to non-approved generative AI tools, treating it as a breach of the confidentiality agreement reviewers sign. Applications: NSF takes a much lighter line, encouraging proposers to indicate in the project description whether and how generative AI was used, and placing responsibility for accuracy and authenticity — including avoiding fabrication, falsification and plagiarism — squarely on the proposer. Check your own funder and, separately, your institution’s acceptable-use policy before you start, because these rules are revised often and are not uniform across agencies.

A language model predicts plausible text. A reference is a highly patterned string — author, year, title, journal, volume — so a model can generate one that is structurally perfect and entirely fictional, or worse, one that mixes a real author with a real journal and a paper neither ever produced. The failure mode is not random noise; it is confident, well-formed and specifically difficult to spot by eye. Retrieval-based tools reduce this because they return records they actually looked up rather than composing them, but they do not eliminate it: a retrieved paper can still be summarised inaccurately. The only reliable defence is mechanical. Resolve every DOI, open every paper you cite, and confirm the sentence you attributed to it is actually in it.

Not for anything that has to be defensible. In a 2025 study in Cochrane Evidence Synthesis and Methods, Lau and Golder compared Elicit against the original traditional searches of four systematic reviews and found average sensitivity of roughly 39-40 per cent, against about 94.5 per cent for the traditional searches. Precision was substantially better — around 42 per cent versus about 7.6 per cent — which tells you exactly what these tools are for: they return a cleaner but much smaller slice. Their own conclusion was that Elicit is not sensitive enough to replace a traditional systematic-review search and is useful as a supplementary or preliminary tool. A separate feasibility study across environmental and life sciences reached a compatible conclusion for data extraction, recommending use as a secondary reviewer or sanity check rather than an autonomous one. Treat these findings as applying to the category, not to one vendor.

Usually yes, and increasingly with specifics. COPE’s position statement on authorship and AI tools holds that AI cannot be an author, and requires authors who use generative AI in writing, in producing images, or in collecting or analysing data to disclose which tool was used and how, in the methods or equivalent section. In evidence synthesis, the joint Cochrane, Campbell Collaboration, JBI and Collaboration for Environmental Evidence position statement of 31 October 2025 goes further: any AI use that makes or suggests a judgement — eligibility, extraction, risk of bias, certainty of evidence — must be disclosed with the system name, version, purpose and known limitations, and authors must describe how they verified the output. Both draw the same practical line: spelling, grammar and structural polish generally does not require disclosure; anything that shapes a finding does.

Anything you do not have the right to disclose to a third party. In practice that means: manuscripts you are reviewing, grant applications you are reviewing, unpublished data from collaborators, identifiable participant data, anything under a confidentiality or material transfer agreement, and invention disclosures you have not yet filed on. The reviewing cases are not a matter of judgement — both NIH and NSF treat uploading review material to a generative AI tool as a confidentiality breach. For everything else, the operative question is not whether the tool is trustworthy but whether you are permitted to hand the material over at all.

Semantic Scholar is free, including a public API, and is built by the Allen Institute for AI. Zotero is free and open source, with paid storage above a quota. ASReview is free and open source for screening. Connected Papers and ResearchRabbit have usable free tiers. Elicit has a free Basic tier with paid tiers above it. The general assistants all have free tiers, but those tiers generally carry the weakest data-handling terms, which matters more in research than the feature difference does. Pricing in this category changes frequently enough that you should check the vendor’s own page rather than any comparison, including this one.

Not neutral: Ask CASRAI is our own product, and you should read this entry accordingly. It is a retrieval-augmented question-answering system over a corpus of CASRAI’s own published pages plus six external feeds — the U.S. Federal Register, a research and grants slice of it, Grants.gov, Regulations.gov, NSF News and UKRI — re-ingested daily. It is specialised for research administration rather than trained on it; it cites a source for every factual claim and answers that it does not have something in the corpus rather than guessing. It is deliberately narrow: it answers questions in short form and refuses to write documents, so it will not draft a proposal, a Specific Aims page or a data management plan, and it does not cover every funder or regulator. It is a paid feature of a Regulatory Radar subscription, with no free tier. For the jobs in the other four columns of the table above — literature discovery, drafting, reference management, systematic review screening — the third-party tools listed there are the right answer and we are not the competitor.