How to conduct a literature review with AI

Lock the question and protocol before searching

Write the review question, population or context, intervention or phenomenon, comparison where relevant, outcomes, study types, date range, languages, databases, and exclusions. Define what the review will not answer. Record the protocol version and any later amendments with reasons.

PRISMA 2020 is a reporting guideline for systematic reviews. Its checklist and flow diagrams support transparent reporting of why a review was done, which methods were used, and what was found. It does not make a search comprehensive by itself, so adapt methods to the discipline and review type.

Create a reproducible search ledger

For every database, save the platform, complete query, filters, search date, result count, and export filename. Ask AI for synonyms and controlled-vocabulary candidates, but test each proposed term. Run known-item checks: can the query retrieve several clearly relevant studies identified before the search?

Use stable record identifiers such as DOI, PMID, accession number, or repository URL. Never accept a model-produced reference until its title, authors, year, and identifier match a real source.

Worked synthetic review

Suppose the question is: “What workplace studies examine whether structured source-verification prompts change error detection in AI-assisted analytical writing?”

The protocol includes empirical workplace or realistic-simulation studies, requires an observable verification behavior, and excludes opinion essays and studies measuring confidence alone. A search ledger might show:

| Source | Query family | Results | Included after full text | | --- | --- | ---: | ---: | | Database A | AI assistance + verification + writing | 84 | 4 | | Database B | generative AI + fact checking + workplace | 61 | 3 | | Registry C | AI writing intervention | 17 | 1 |

These counts are synthetic and do not describe an actual evidence base.

Screen with dual visibility

Keep one row per record with duplicate status, title-screen decision, full-text decision, exclusion reason, reviewer, and date. AI may propose a decision and quote the relevant passage, but a human reviewer should verify the record under the agreed process. Sample both included and excluded items for consistency.

A bad exclusion says “not relevant.” A useful exclusion says “No observable verification behavior; outcome is self-reported confidence only.” Specific reasons make errors auditable and feed the flow diagram.

Extract claims, not impressions

Use a study table for design, participants, setting, intervention, comparator, outcome definition, result, uncertainty, limitations, and funding. Bind every extracted value to a page, table, figure, or supplement location. If the full text is unavailable, do not infer missing methods from the abstract.

Ask AI to compare tables for inconsistent denominators or outcome definitions. Then verify flagged items in the source.

Synthesize without counting headlines

Group studies by question-relevant mechanism or design, not simply by positive and negative result. Explain differences in population, task, intervention, and measurement. One large observational study and one small experiment are not two equivalent votes.

For the synthetic review, a cautious synthesis might say: “Eight studies met the protocol, but their verification prompts and error measures differ. The collection supports several candidate mechanisms for further testing; it does not establish one transferable effect.”

Verification checklist

Confirm that the protocol predates screening; every database query is preserved; duplicates remain traceable; exclusions have reasons; each citation resolves to the original record; extracted claims match the source; amendments are logged; disagreement is visible; and conclusions stay within the reviewed designs.

Sources and claim map

The question, databases, counts, and synthesis are synthetic. This educational workflow does not establish that a review is exhaustive, clinically valid, or suitable for a consequential decision without domain review.

• PRISMA, PRISMA 2020: supports transparent reporting through the statement, checklist, expanded checklist, and flow diagrams.
• PRISMA, PRISMA-S search reporting resources: supports reporting literature-search methods and record flow.

Sources: [1] [2]