Checking prior work
Is there an AI tool that checks if my research question has already been studied?
Updated by the Vidi team
Plans start at $5 a month.
You have a research question and you want to know two things. Has someone already studied it, and which papers come closest? AI search tools can help with both. The better ones search by meaning, so they can find papers that describe your idea in different words.
They cannot give you a yes or no on novelty. Each tool searches its own index, ranks what it finds and stops. This guide explains what the main tools do, how far you can trust them, and a routine that combines AI search with citation chasing and keyword searching. That way you can say more confidently whether your question is still open.
How it works
How to check your question with Vidi Research
Ask your question in plain words
On Vidi Research, write your research question the way you would say it to a supervisor. Vidi turns it into searches across OpenAlex and arXiv. Asking is free with a Vidi account, up to 20 questions a day. Questions work in English.
Read the closest papers and the novelty score
You get the three closest papers, each with a relevance score, a one-line reason and a quote copied from its abstract. A novelty score from 0 to 100 shows which parts of your question look covered and which look open. A graph shows up to 24 related papers linked by real citations. Open each paper and check it yourself.
Widen the search, then read and review
Use the best matches as seeds for citation chasing and a keyword search in another database. Add papers to your library and read them with Vidi. You can also draft a literature review from 2 to 12 papers for 5 credits, or turn one paper into a short narrated video.
What an AI search can and cannot tell you
An AI search tool shows you the closest papers it finds in its own database. It cannot show you papers it never indexed, and it cannot tell you that nobody has asked your question. One methods guide suggests turning the task around. Try hard to prove your question has already been answered, and call it new only if it remains meaningfully unresolved after that effort.
The same guide separates two ideas. A question can be previously studied, meaning the general or a closely related question has been investigated, without being answered. Say which one you mean when you write your proposal.
Also watch for overlap between tools. Elicit and Undermind both draw on Semantic Scholar, so running both gives you less independent coverage than it seems. One recent paper names a risk worth remembering, the 101st-paper trap: after you read the first 100 papers in a field, the idea that comes next may already be in the 101st.
A search routine that holds up
Break your question into parts: the concepts, the relationship you care about, the population, the context, the mechanism and the outcome. List synonyms for each part, including the terms other fields use. Keyword search depends heavily on the exact words you choose.
Then combine three kinds of search. Start with an AI search by meaning. Take the best two or three papers it returns and use them as seeds in a citation map such as ResearchRabbit, Connected Papers or Litmaps, or follow their references and citations by hand. Finish with a keyword search in Google Scholar or a specialist database such as PubMed, Scopus or IEEE. Look for review articles on purpose, and check preprints and the most recent months separately, because the newest work can be missing.
Open and read the closest papers yourself. Cite the original papers, not an AI-generated report. Keep a log of the date, the tool, your prompts and what came back. AI tools often do not show their underlying query, so your notes may be the only record of the search.
How to judge whether a tool covers your field
Recall varies. In an independent study of Elicit published in Cochrane Evidence Synthesis and Methods, its search found 36 of 52 relevant studies in one case study and only 16 of 45 in another. It also found some studies the original review authors had missed. A systematic review of Elicit called it a complementary tool.
Vendor figures are harder to weigh. Undermind's own whitepaper claims very high accuracy in classifying papers. Treat numbers like this as claims, not independent evidence.
You can test a tool yourself. Ask it a question you already know well and see whether the papers you expect come back. For a firmer test, take a published systematic review in your area, run its inclusion question as a search, and count how many of the review's included studies the tool finds. If it misses many of them, do not rely on it alone for your own question.
Common mistakes to avoid
Taking no close results as proof that the question is new. An empty result only tells you about that tool and that wording.
Searching with one phrasing and missing the terms another field uses for the same idea.
Running several tools that sit on the same underlying database and counting them as independent checks.
Trusting a gap-finder's suggested gaps without looking for a newer paper that already fills them. One review of these tools notes that tools which do not search for newer studies can report gaps that are already closed, and that AI-generated references can be wrong.
Assuming the tool read full papers when it may have used only titles and abstracts. Assuming good coverage of humanities and non-English work, where reviewers have found gaps.
Using AI search alone where you need a search others can reproduce, such as a systematic review.
Ways to find the papers closest to your research question
| You start with | Good for | Watch out for | |
|---|---|---|---|
| Elicit | A question in plain language | Papers that match the meaning of your question; pulling data into tables, strongest in biomedical, psychological and social science topics | One independent study found 36 of 52 relevant studies in one case and 16 of 45 in another; its underlying search query is not visible |
| Undermind | A detailed question, often with several parts | Repeated searching that estimates what share of the relevant papers it has found so far | No reproducible search strategy; sources disagree on full text versus abstracts; the 2024 free tier was ten searches a month |
| Consensus | A focused question | A quick read on whether the evidence leans yes, no, possibly or mixed | It will not show all the differences between the studies |
| Citation maps (ResearchRabbit, Connected Papers, Litmaps) | One or more seed papers you trust | Related papers around your seeds; Litmaps shows how a set of papers interact over time | You need good seeds first, and it cannot prove you found every relevant paper |
| scite | A specific finding or paper | Showing whether later research supported or contradicted a finding | Not designed for finding the closest papers to a research question |
| Vidi Research | A question written as you would say it to a supervisor | Three closest papers with abstract quotes and relevance scores, a 0 to 100 novelty score and a citation graph of up to 24 papers | Searches OpenAlex and arXiv; three papers is a short list; questions in English only |
| Google Scholar and PubMed | Keywords; PubMed also supports MeSH subject headings | Broad coverage: about 400 million citations in Google Scholar and 37 million in PubMed, by one 2025 review's figures | Results depend heavily on the exact keywords you use |
Limits worth knowing
- No AI tool can prove a question is new. Each searches its own database, and an independent test of one tool found it missed a large share of relevant studies.
- Vidi Research searches OpenAlex and arXiv and shows the three closest papers, so treat it as a starting point and confirm with other databases and citation chasing.
- The novelty score and abstract quotes show you where to look, not what the papers found. AI output can be wrong, so read the papers before you rely on them.
- Questions on Vidi Research work in English only, and AI search on its own does not give you a reproducible strategy for a systematic review.
Sources
- How to check whether AI says your research question has already been answered
- Paper describing the 101st-paper trap
- Lau et al. 2025, Elicit search evaluation, Cochrane Evidence Synthesis and Methods
- Elicit as an AI-Assisted Research Tool: A Systematic Review
- University of Northampton library guide: Elicit
- Product review of AI search tools including Undermind (PMC)
- Katina: Undermind AI shows the power of successive search
- CASRAI guide to Undermind
- Tooliverse: Undermind
- Undermind founders on Hacker News
- Undermind whitepaper
- Semantic Scholar paper (arXiv)
- Best AI tools for academic research (thesify)
- AI tools for academic research (Atlas)
- Connected Papers overview (ResearchPal)
- scite
- Research gap finder tools (Paperguide)
- Systematic literature review with Elicit AI: use cases and limitations
- Elicit AI review (aifocus)
- Elicit vs Consensus (thedrive)
- Best AI research tools for beginners
If the AI finds nothing close, can I say my question is new?
No. An empty result only means that tool, with that wording, found nothing in its own database. Search again with synonyms, chase citations from the nearest papers, and run a keyword search in Google Scholar or a specialist database. One guide suggests calling a question new only if it remains meaningfully unresolved after a real effort to prove it has been answered.
Which tool should I start with?
For a detailed question with several parts, Undermind is built to keep searching and estimate how much it has found. Elicit suits empirical fields such as biomedical, psychological and social science topics. Consensus gives a quick read on whether evidence leans yes or no. If you already have one good paper, a citation map is a useful next step.
Do these tools read the full papers?
Not always. Two sources say Undermind's classification step uses titles and abstracts, while another says it reads full-text PDFs, so check the tool's current documentation. Either way, a match on an abstract tells you a paper is close to your question, not what it found. Read the closest papers yourself.
Can I report an AI search in my thesis or systematic review?
Not as your only search. One review calls the lack of a reproducible search strategy Undermind's weak point and says most AI search platforms share it. A study of Elicit found its underlying query opaque. Use database searches you can document, and log the date, tool and prompts of any AI search you add.
Are Elicit and Undermind two independent checks?
Less than they seem. Both draw on the Semantic Scholar corpus, so their results overlap. For a more independent check, add a citation map built from your best papers and a keyword search in a different database, such as Google Scholar, PubMed, Scopus or IEEE.
Is Vidi Research free to use for this?
Asking questions is free with a Vidi account, up to 20 a day, and no credit card is needed. Each answer gives the three closest papers, a novelty score and a citation graph. A literature review costs 5 credits and a paper video costs 9 credits. New accounts start with 5 free credits.