How to use AI for a literature review without losing the evidence

AI can reduce repetitive literature-review work, but it does not remove the need for a defined method, transparent searches, human judgement and source-level verification.

In brief

A useful AI-assisted literature review begins with the review method, not the model. A narrative review, scoping review and systematic review answer different questions and require different levels of search coverage, screening documentation and reporting.

Use AI for bounded tasks such as generating candidate search terms, prioritising records, extracting structured fields and comparing findings. Keep the protocol, source records and final decisions outside the model response so another person can inspect how the review was produced.

Define the review before choosing a tool

Write down the question, audience, review type, date range, languages, eligible study designs and intended output before searching. For a systematic review, predefine inclusion and exclusion criteria and decide how disagreements will be resolved.

This boundary prevents the review from quietly changing when an AI system suggests an interesting but out-of-scope paper. It also makes tool selection easier: a broad narrative overview needs different controls from a review intended to support a clinical or policy decision.

State the research question in plain language and, where appropriate, a structured framework such as PICO.
Name the review type and the reporting guidance that applies.
Separate exploratory searches from the searches that will be reported as part of the review method.

Build a search that can be explained

AI can suggest synonyms, related concepts and spelling variants, but every proposed term still needs subject knowledge and a check against known relevant papers. Translate the accepted concepts into the syntax of each database rather than sending one unchanged prompt everywhere.

Record the database or index, interface, full query, filters and search date. Semantic search and citation-network exploration can complement this process, but they should be labelled separately because their ranking and coverage differ from a reproducible Boolean search.

Test whether known relevant studies appear in the result set.
Search more than one appropriate source when review completeness matters.
Preserve each retrieval run instead of retaining only the final combined library.

Give AI bounded, reviewable tasks

The safest unit of automation is a task with a visible input, an explicit output format and a human checkpoint. Asking for a final review in one step hides where papers were missed, fields were inferred or conflicting evidence was flattened.

For screening, require a proposed decision, the criterion used and the text that supports it. For extraction, distinguish values copied from a source from interpretations made by the model. Uncertainty should remain visible rather than being converted into a confident answer.

Use AI to propose, prioritise or extract; keep consequential decisions reviewable.
Require supporting passages for study characteristics and findings.
Escalate ambiguous records instead of forcing a binary classification.

Keep screening and extraction auditable

Deduplicate records without discarding provenance: two database records may describe the same work, while a preprint and journal article may be related but materially different versions. Preserve the identifiers and retrieval sources that explain how the consolidated record was formed.

During screening, retain exclusion reasons at the stage where the decision was made. During extraction, save the source location, reviewer status and any correction. These records are more useful than a chat transcript because they connect each decision to the study it affected.

Retain DOI, database identifiers and source URLs where available.
Record human overrides of AI suggestions and the reason for each override.
Version extraction forms when the requested fields change.

Synthesize claims, not just summaries

A literature review should explain patterns, differences and uncertainty across studies. Paper-by-paper summaries can be a useful preparation step, but they do not establish whether methods, populations and outcomes are comparable.

Create a claim-to-evidence table before drafting. For each conclusion, identify the supporting and conflicting studies, relevant limitations and the strength of the inference. Then verify every citation against the original source before it enters the manuscript.

Group evidence by question, theme, method or outcome rather than by document order.
Keep conflicting results visible instead of averaging them into a generic statement.
Mark where a sentence is interpretation rather than a directly reported finding.

Report the AI contribution and its limits

PRISMA and PRISMA-S are reporting frameworks, not certifications for an AI product. A tool can support a PRISMA-oriented workflow without making the resulting review compliant. The authors remain responsible for applying the relevant checklist to the completed work.

Document which tool and version supported each stage, what information it received, how outputs were checked and where its limitations could affect the result. This makes the assistance inspectable and helps readers distinguish automated support from author judgement.

Disclose material AI use according to the rules of the target institution, funder and journal.
Describe human verification and any validation performed within the review.
Do not present coverage, recall or accuracy as proven unless it was measured for the relevant task and dataset.

Use it in practice

Practical AI-assisted literature review checklist

01Define the question, review type, scope and eligibility criteria before retrieval.
02Select databases and supplementary search methods appropriate to the discipline.
03Save exact queries, filters, interfaces and search dates.
04Check candidate search terms against known relevant and irrelevant papers.
05Require source passages and uncertainty labels for AI-assisted screening or extraction.
06Record duplicate handling, exclusion reasons, reviewer changes and human overrides.
07Verify every manuscript claim and citation against the correct source version.
08Report the AI tasks, tool versions, checks and limitations in the final method or disclosure.