RESEARCH / BLOG

Research,
Explained.

Read the research question, follow the authors’ methods and explore the evidence behind the article.

Research Blog

Research methodology · Research insights

AI-Assisted Literature Search: Find Research Gaps Without Invented References

Use AI for search terms, then verify scholarly records, DOIs and source claims yourself. Build a reliable literature review and evidence-based research-gap statement.

By NPT Research6 min read
The novelty toolkit · S206 / 09
A convincing citation is not enough.Check the record. Read the evidence. Refine the search.
NPT Research

The Novelty Toolkit · Series 2 · Part 6 of 9

An AI assistant can produce a beautiful list of ten papers in seconds. But imagine a librarian who can recall convincing book titles and occasionally invents one. You would never cite a book without checking that it exists. Scientific literature searching needs the same discipline—especially when AI is involved.

The safest workflow is AI proposes search strategies; scholarly databases return real records; the researcher verifies what those records establish.

Begin with a human-built search brief

Write the subject, the proposed mechanism, the comparison, the environment and the question you need the literature to answer. Example: 'I need to identify studies of temperature-associated measurement drift in low-cost water-quality sensors deployed for extended periods outdoors.'

Do not ask 'Find a novel idea.' Ask which existing methods, measurement problems and validation approaches the literature has documented. Broad prior-work understanding comes before a gap claim.

Use AI to expand vocabulary

A method may have been described as 'measurement drift', 'instrument stability', 'calibration transfer' or 'ageing response'. AI can propose alternative technical phrases that would not have occurred to you. It should label terms as suggestions, not proven equivalents.

Example prompt

My human-defined research question is: [question]. Generate a search-term matrix with four groups: phenomenon, instrument or population, mechanism or method, and conditions. Include synonyms, broader terms and plausible older terminology. Flag ambiguous terms that may produce irrelevant hits. Do not invent publications or claim that the search has already been completed.

Convert the output into Boolean searches and test the terms against real retrieved abstracts. Remove irrelevant synonyms; add terms discovered in high-quality papers.

The novelty toolkit · Series 2Give each part of the search a clear owner.The assistant proposes terms; real sources establish what is known.
  1. 01
    Researcher briefDefine the question and search scope.
  2. 02
    AI vocabularySuggest synonyms and flag ambiguity.
  3. 03
    Database retrievalRun and record actual searches.
  4. 04
    Source verificationMatch metadata and inspect the relevant evidence.
  5. 05
    Bounded gapDescribe uncertainty within the search coverage.
A learning framework, not a measured result. Verify evidence before making a research claim.
Figure 1. AI improves vocabulary, while databases and human readers establish the evidence base.

Execute the searches yourself

Search Google Scholar and at least one relevant subject-specific database. Use IEEE Xplore or suitable engineering databases for hardware and algorithms; PubMed for relevant biomedical questions; additional indexing tools where available. OpenAlex offers searchable records and citation metadata through its API (OpenAlex’s API documentation).

Record the database, search date, exact query and filters. Do not paste the same syntax indiscriminately into every platform. A date-limited AI search may miss older foundational solutions; a non-English or preprint source may use unfamiliar terminology.

Turn AI into a retrieval auditor

Once you have actual titles and DOIs, use AI to check which search concepts each paper appears to cover, or to suggest other paper families worth searching. But do not assume an AI can verify citations merely because it outputs a DOI. Open the DOI, inspect the publisher record, check the author/title/year match and examine the full text where relevant.

Example prompt

I will provide verified publication records. Do not create additional citations. For each record, identify which components of my question it might address based only on the supplied title and abstract. Mark any conclusion requiring full-text verification. Suggest the next five search queries, and explain the gap each query is designed to test.

Keep an evidence-status ledger

Use four columns: AI suggestion, database result, verified source evidence, remaining uncertainty. If an AI mentions a paper that you cannot locate in the library catalog, Crossref, publisher site or DOI system, mark it unverified or reject it. Do not silently replace it with a vaguely similar paper.

Research on LLM hallucinations documents that confident language can coexist with unsupported statements (Farquhar and colleagues’ semantic-entropy study). The key response is verification, not simply asking the same model whether it is certain.

The novelty toolkit · Series 2Follow a suggestion all the way to its evidence.A DOI that resolves is only the first check.
  1. 01AI suggestionTreat the proposed term or citation as unverified.
  2. 02Retrieved recordMatch title, authors, date and identifier.
  3. 03Source passageCheck the method, comparison and reported outcome.
  4. 04UncertaintyRecord what the source still does not establish.
A learning framework, not a measured result. Verify evidence before making a research claim.
Figure 2. The quality of a gap claim depends on real retrieval and independently checked evidence.

The 'no results' trap

Suppose the exact phrase 'autonomous adaptive field-drift intelligent sensor' returns nothing. It is tempting to celebrate. Instead search the components independently, use synonyms, inspect references and patents, and look for methods hidden in other application domains. No exact phrase match is evidence about wording, not a proof of scientific originality.

A defensible interim statement describes the databases and dates searched and specifies which question was not answered by the retrieved studies. If you have only completed an exploratory search, say so. PRISMA-S provides a useful reference for transparent search reporting in systematic reviews but does not turn a quick scan into a systematic review (the PRISMA-S search-reporting guideline).

Practice: verify every citation

Ask an AI tool for search terms, not paper titles. Use those terms in a scholarly database. Choose three real papers, verify their citation metadata, and read the relevant parts. Then ask AI for two alternative search directions based on those papers. Record any new relevant source that changes your gap claim. The purpose of the exercise is better coverage, not a longer bibliography.

Example of a citation-verification decision

Imagine an AI assistant supplies an article title, three authors, a year and a DOI. You open the DOI and discover a different title. Do not silently cite the first one because the DOI looks formal. Record the mismatch and look for the actual paper through a publisher site, a cross-disciplinary index or an institutional library.

Next imagine the paper exists, but the AI says it proves that adaptive calibration eliminates field drift. Read the methods and results: perhaps the paper tested only one instrument in the laboratory and reported a partial reduction. The correct research note is much narrower. Verification therefore has two stages: metadata authenticity and accuracy of the scientific claim.

What different sources contribute

A journal article may report a validated finding; a preprint may reveal a recent but not yet peer-reviewed result; a patent publication may disclose a technical concept; a thesis may describe a method in more detail than a short article; a trial registry may reveal that a study is underway. None should automatically be treated as equivalent evidence of effectiveness.

The source type matters to interpretation. A patent can threaten a patentability claim without showing that a device works. A registered protocol can show a question is being studied without demonstrating its answer. A preprint can be timely and useful but needs careful reading of validation and revision status.

A five-check citation audit

Check whether each record actually exists; whether title, author, year and DOI match; whether the claimed method appears in the relevant full-text section; whether the numerical outcome has the correct comparator; and whether later or corrected publications change the conclusion. Make this audit part of your research notebook, not a final-minute reference-cleaning task.

What to remember

The AI assistant should not determine what the literature contains from conversational memory. It should help you find, organize and question the actual literature. The next tutorial shows how to compare that literature without letting a convenient summary hide important methodological differences.

Methodological context and limits

Farquhar and colleagues’ semantic-entropy study evaluates a method for detecting a particular subset of model hallucinations. It does not make ordinary chatbot citations trustworthy. Cochrane’s study-search guidance remains useful for planning the actual evidence search and its documentation.

Continue the research novelty series

Series 1: originality, evidence and scientific judgment

Begin with what makes research novel, then find the closest prior research and build a novelty matrix. Before using an assistant, examine the limits of AI novelty judgments and why the researcher owns the question. Develop the conversation with scientific prompting and six-round brainstorming, then challenge the idea against evidence. The complete worked research dialogue connects those decisions.

Series 2: from a problem to a defensible proposal

First separate a research gap from a research problem and identify a genuine problem worth investigating. Next search published literature for the gap and align the question, objectives and hypothesis. For AI-assisted work, clarify the problem without outsourcing judgment, verify the literature search and its references and compare papers in an evidence matrix. Then stress-test the research hypothesis before following the complete observation-to-proposal workflow.

Sources and further reading

  1. Rethlefsen et al. PRISMA-S: an extension to the PRISMA Statement for Reporting Literature Searches in Systematic Reviews. Systematic Reviews (2021).
  2. Cochrane Handbook for Systematic Reviews of Interventions. Chapter 4: Searching for and selecting studies.
  3. OpenAlex. API reference: scholarly works, search and citation metadata. Official documentation.
  4. Farquhar et al. Detecting hallucinations in large language models using semantic entropy. Nature 630, 625–630 (2024). Published 19 Jun 2024