Using AI for a Literature Review: Workflow, Verification and Reporting

Searching “literature review artificial intelligence” usually means one thing: how to use AI tools inside a literature review without damaging its validity. (If you instead need to write a review of the AI field, the structure advice in layout of a literature review applies unchanged.) This guide is for students, researchers and librarians who want to know which stages AI can help with, what must still be checked by a person, and what to record so the review can be reported honestly.

Purpose: what AI can and cannot do in a literature review

AI tools can speed up searching, summarising, citation management and thematic analysis, but they cannot judge research quality, and a summary is a starting point rather than a substitute for reading the paper. That is the position taken by the IADT Library’s guide to AI and literature reviews, which also lists the main limitations: incomplete coverage (paywalled sources are missed), unreliable search results, invented references, and text prediction rather than evidence-based reasoning.

For formal evidence synthesis the bar is higher. In November 2025 Cochrane, the Campbell Collaboration, JBI and the Collaboration for Environmental Evidence issued a joint position statement, published in four journals, supporting the RAISE framework (Responsible use of AI in evidence SynthEsis). It asks synthesists to take responsibility for any AI they use and to make sure it “will not compromise the methodological rigour or integrity of their synthesis”, with human oversight, transparency and justification, and open reporting of AI use.

Workflow by stage, with what a person must still check

The table below is an editorial workflow. The checks in the right-hand column follow the limitations and requirements in the sources listed at the end; the stage names are a common review sequence, not a standard.

Stage	Where AI may help	What a person must still do
1. Scope the question	Suggest related terms and seed topics	Decide the question and eligibility yourself; keep your own insight central
2. Search	Surface candidate papers and related work	Run and record structured searches in library databases; seed with papers you found there, because AI tools can miss paywalled sources and return unreliable results
3. Screen	Rank or pre-sort records	Decide inclusion; for a systematic review, record how many reviewers screened each record and any automation tool used (PRISMA 2020 item 8)
4. Extract data	Pull draft values from papers	Check every extracted value against the paper; record automation used (PRISMA 2020 item 9)
5. Appraise quality	Little or none	Judge quality yourself: AI cannot judge research quality; record automation if used (PRISMA 2020 item 11)
6. Synthesise and write	Suggest themes or an outline	Read the sources behind every theme; write the argument yourself and disclose AI assistance where required

Verifying every reference

AI can invent references that look real. Before any citation enters your draft, confirm that the paper exists and says what you claim.

1. Look the title up in a library database or Google Scholar and open the actual record.
2. Match authors, year, journal and DOI against the record, not against the AI output.
3. Read the part of the paper that supports your claim; do not rely on the AI summary.
4. Delete any reference you cannot locate.

This is the verification step the IADT guide recommends (check references against library databases or Google Scholar). A tool that summarises a retrieved paper is working from real text; one that generates a citation from memory may not be, which is why the existence check comes first. Our Consensus guide and the Elicit vs Consensus comparison describe how two such tools present their sources.

What to report when you used AI

PRISMA 2020 asks for “details of automation tools used in the process” under item 8 (selection process), item 9 (data collection process) and item 11 (risk of bias assessment). It also asks for the full search strategies for all databases, registers and websites, including filters and limits (item 7), and every source searched with the date it was last searched (item 6). If you use an AI tool at any of these points, those items are where it is reported. The RAISE-aligned joint statement adds the expectation of transparent reporting and justification of AI use.

Whether you also need a separate disclosure statement depends on your institution, course or target journal. Confirm that AI use is permitted for your coursework or submission before you start. For wording models, see AI disclosure statement examples, and for one publisher’s rules see the Elsevier generative AI policy.

A short annotated example: an AI-use log

Illustrative template, not a real review. Keeping a log as you go makes the reporting items above easy to complete.

Stage	Tool and date	What it was used for	Human check done
Search	[tool name, version, date]	Find candidate papers from three seed articles	All candidates re-found in a library database; search strings saved
Extraction	[tool name, version, date]	Draft a table of sample sizes	Each value checked against the paper by one reviewer

The annotation to copy is the last column: every AI step has a named human check beside it.

Common mistakes

• Citing a reference taken from an AI answer without opening the paper.
• Treating an AI search as the whole search, so paywalled or poorly indexed work is silently missing.
• Using an AI summary in place of reading the source.
• Delegating quality appraisal to a tool that cannot judge research quality.
• Not recording which tool was used, when, and for what, so PRISMA items 8, 9 and 11 cannot be completed.
• Assuming AI use is allowed without checking the course, institution or journal rules.

Checklist

• AI use is permitted for this assignment or submission.
• Seed papers came from library databases.
• Full search strategies, databases and last-searched dates are saved.
• Every reference exists and was checked against a database record.
• Every AI-extracted value was checked against the paper.
• Quality appraisal was done by a person.
• AI tools, dates and purposes are logged and reported.

FAQ

Can I use AI for a literature review?

Often yes for support tasks such as searching and summarising, provided the rules of your course, institution or journal allow it and you verify the output. The IADT guide advises confirming that AI use is permitted for coursework.

Can AI replace reading the papers?

No. Summaries are a starting point, and AI cannot judge research quality.

Does PRISMA 2020 require me to report AI tools?

It asks for details of automation tools used in the selection, data collection and risk-of-bias steps, where applicable.

Is a narrative review held to the same reporting standard as a systematic review?

Not necessarily; see narrative review vs systematic review. The verification steps above apply to both.

Sources

• Cochrane, Campbell Collaboration, JBI and Collaboration for Environmental Evidence, joint position statement on responsible AI use in evidence synthesis, 11 November 2025: cochrane.org (read 2026-11-08).
• PRISMA 2020 checklist, items 6-9 and 11: prisma-statement.org (read 2026-11-08).
• IADT Library, AI and Literature Reviews: iadt.libguides.com (read 2026-11-08).