4. Research, Literature Discovery, and Evidence Checking

AI can help researchers move from a broad topic to a manageable question, identify search concepts, map debates, compare papers, and organise notes. It must not replace systematic searching, critical appraisal, or direct engagement with original sources.

4.1 Research discovery tools

Scholarly search engines and databases remain essential for discovering academic literature. Google Scholar, PubMed, Crossref, institutional library databases, Scopus, and Web of Science serve different coverage and indexing purposes. AI-assisted discovery tools such as Elicit, Consensus, Semantic Scholar, and Perplexity may help identify candidate studies, extract study details, or locate source-linked explanations. These services differ in corpus, coverage, and methods; no single tool should be assumed to provide a complete systematic search.

4.2 A responsible literature workflow

1. Translate the topic into a focused research question and define key concepts, population, setting, and timeframe.
2. Draft search strings with synonyms, spelling variants, Boolean operators, and relevant subject terms.
3. Search more than one suitable database and record the date, platform, query, filters, and number of results.
4. Use AI tools to suggest related concepts, terminology, and candidate papers, but retrieve and inspect the original papers.
5. Screen titles and abstracts using pre-established inclusion and exclusion criteria.
6. Read full texts and extract study design, sample, context, measures, findings, limitations, and relevance.
7. Check each reference in the publisher record, DOI registry, database, or library catalogue.
8. Synthesise patterns, contradictions, methods, and gaps in your own analytical framework.

4.3 AI for reading and synthesis

When using an AI tool to summarise a paper, provide the paper itself and ask for a structured extraction: research question, theoretical lens, design, sample, data collection, analysis, findings, limitations, and relevance to the review. Request quotations only with page numbers and verify every quotation against the PDF. Distinguish the authors’ findings from the AI’s interpretation. When synthesising several papers, ask for a comparison table and then check each row against the source texts.

4.4 Systematic and scoping reviews

AI may support search-term development, deduplication, abstract screening, extraction, and drafting of a review protocol. However, the review team remains responsible for transparent methods, reproducibility, and bias control. Record the tool, version where possible, prompts, dates, screening rules, and human decisions. Do not report AI-assisted screening as fully automated unless that accurately describes the process and has been validated. Follow the relevant reporting standard for the review type.

4.5 Research questions to ask an AI assistant

Useful requests include: “Suggest alternative search terms for this research question”; “Identify assumptions and ambiguities in my inclusion criteria”; “Create a data-extraction template without inventing study findings”; and “Compare these supplied abstracts, marking missing information as not reported.” The instruction to mark missing information rather than infer it is especially important.

4.6 The citation rule

Never copy a reference merely because an AI system supplied it. Confirm that the source exists, the authors and year are correct, the title and journal are accurate, the DOI resolves, and the cited source supports the statement. Research has documented fabricated and inaccurate AI-generated bibliographic references (Fabrication and errors in the bibliographic citations generated by ChatGPT, 2023; Picazo-Sanchez & Ortiz-Martin, 2026). A reference manager such as Zotero, EndNote, or Mendeley can help organise verified records, but it cannot by itself guarantee that a source is relevant or accurately represented.