Suite 4.12, Ashworth House, 18 Kingsley Road, London EC2A 4NE, United KingdomLondon, United Kingdom
  1. Home
  2. Blog
  3. AI in Literature Reviews: A Practical and Ethical Framework

AI in Literature Reviews: A Practical and Ethical Framework

AI tools can screen thousands of abstracts in minutes, and they can also invent sources. A stage-by-stage guide to where AI helps a literature review and where a human must stay in charge.

AI in Literature Reviews: A Practical and Ethical Framework

A thorough literature review used to mean weeks in databases, spreadsheets of abstracts and stacks of printed papers. AI tools now promise to compress that work into days. Some of the promise is real. Some of it can damage a researcher's reputation.

This article sets out a simple framework for using AI in a literature review, stage by stage, with one principle throughout: AI can assist with the work, but the researcher remains responsible for every claim and every citation.

Know what the tools actually are

It helps to separate three kinds of tool that are often lumped together:

TypeWhat it doesMain risk
General chat assistantsGenerate fluent text from patterns learned in trainingMay produce plausible but non-existent references
Search-grounded research assistantsRetrieve real papers from indexes, then summarise themCoverage gaps and over-simplified summaries
Screening and extraction softwareRank or classify records using machine learningMissed relevant studies if used without checks

A general chat assistant is not a database. Asked for "ten key papers on X", it may write citations with real-sounding authors, journals and years that do not exist. Studies testing early versions of ChatGPT found a substantial share of generated references were fabricated or contained serious errors. Newer tools are better, but the risk has not disappeared.

A stage-by-stage framework

Stage 1: Scoping the question (AI is useful)

Brainstorming synonyms, related concepts and possible sub-questions is a good fit for AI. Ask it to suggest alternative terms for your key concepts, then check those terms against controlled vocabularies such as MeSH.

Stage 2: Building the search (AI helps, you verify)

AI can draft Boolean search strings for PubMed, Scopus or Web of Science. Always test them yourself: run the search, check that known key papers appear, and adjust. Record the final strings exactly; they belong in your methods section.

Stage 3: Screening (AI can prioritise, humans decide)

Machine-learning screening tools can rank thousands of titles and abstracts so the most likely relevant ones come first. This saves real time. For a systematic review, however, inclusion decisions should still be made or confirmed by human reviewers, ideally two working independently, and the use of the tool should be reported in line with PRISMA 2020.

Stage 4: Reading and extraction (be careful)

Summaries generated by AI can drop important qualifiers, merge findings from different studies or misstate sample sizes. Use them to decide what to read, not as a replacement for reading. Extract data from the original paper, not from a summary of it.

Stage 5: Synthesis and writing (you write it)

The argument, the interpretation and the judgment about quality of evidence are the intellectual contribution of the review. AI can help you tighten sentences or check grammar. It should not generate the substance.

The golden rule for references: never cite a paper you have not opened. Check every DOI, author list and year against the publisher's page or a trusted index before it enters your reference list.

Integrity and disclosure

Publishers and bodies such as COPE (the Committee on Publication Ethics) agree on some basics. AI tools cannot be listed as authors, because they cannot take responsibility for the work. Authors who use AI in preparing a manuscript should say so, usually in the methods or acknowledgements, describing which tool was used and for what.

A short disclosure might read: "An AI assistant was used to suggest search-term synonyms and to improve the language of the introduction. All searches, screening decisions, data extraction and interpretation were carried out by the authors, who take full responsibility for the content."

Protect your data

Do not paste unpublished manuscripts, confidential peer-review material, or personal data from participants into public AI tools. Many services may retain inputs. In the UK, personal data is protected under UK GDPR and the Data Protection Act 2018, and your institution's policies will usually be stricter than a tool's terms of service.

A balanced view

Used well, AI shortens the dull parts of a literature review and leaves more time for thinking. Used badly, it produces confident errors that reviewers and readers will eventually find. The difference lies in verification, disclosure and the researcher's own judgment.

Our workshop Generative AI in Academia: Responsible Use covers these practices in hands-on sessions.

References

  1. Page, M. J. et al. (2021). The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ, 372, n71.
  2. Walters, W. H. & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, 14045.
  3. Committee on Publication Ethics (2023). Authorship and AI tools: COPE position statement.
  4. Information Commissioner's Office. Guide to the UK General Data Protection Regulation (UK GDPR).
KSR Press Editorial Team

Articles from the editors and tutors at KEYSTONE Research Press on research methods, publishing and AI.