How to Use AI for a Literature Review (Without Getting Burned)
AI tools can take weeks off a literature review. They can also hand you confident summaries of papers that do not exist. The difference is the workflow. This guide walks through where AI genuinely helps, where it fails, and how to use it in a way you can defend to a supervisor or reviewer.
Where AI helps
- Scoping. Turning a vague topic into a set of search terms, synonyms and related concepts.
- Triage. Reading 200 abstracts and sorting them into “relevant”, “maybe” and “no”, with a one-line reason for each.
- Extraction. Pulling the sample size, method, setting and main finding out of each paper into a table.
- Questioning a paper. Asking a specific paper what its limitations are, or whether it controlled for something, and getting an answer with the page it came from.
- Finding gaps. Asking what the collected papers disagree on, or what none of them measured.
Where AI fails
- Finding papers from memory. A general chatbot asked for “five studies on X” will often invent authors, years and journals. Search in a database, then give the AI the real papers.
- Judging quality. It cannot tell a preprint from a replicated finding unless you ask it to check specific things.
- Writing the synthesis. It will produce a smooth paragraph that reads like a review, with the argument flattened. The synthesis is the part that is yours.
A workflow that holds up
1. Search in real databases first
Use Google Scholar, Scopus, Web of Science, PubMed or your field’s database to build the candidate set. Export the records. Our guide to research databases covers the main ones. If you are doing a systematic or scoping review, follow the protocol for that type; see how to write a systematic literature review.
2. Screen with AI, decide yourself
Give the tool your inclusion criteria and the abstracts, and ask for a decision with a reason per paper. Spot-check a sample, and read every “maybe” yourself. Keep the AI’s reasons; they are useful when you write up the screening.
3. Read the PDFs with AI beside you
Upload the papers you kept and ask questions of each one: what was measured, how, on whom, and what the authors say the limitations are. Insist on page-level citations so you can check. This is the step where a tool built for research, such as Avidnote, is different from a general chatbot: the answers come from the PDF in front of you, with the page cited, and your notes stay attached to the paper.
4. Build the extraction table
One row per paper, columns for the things your review compares. Let the AI fill a first draft of each row from the PDF, then correct it. The table becomes the backbone of your synthesis and your appendix.
5. Write the synthesis yourself
Use AI to check your draft, not to write it: ask it to find claims without a citation, to point out where two paragraphs say the same thing, or to list papers in your table that your text never mentions. See our basic guide to writing a literature review for structure.
Keep a record
Note which tool and model you used, for which steps, and on which dates. Save the prompts you reused. Most journals and many universities now ask for an AI-use statement; our disclosure statement generator builds one in the publisher’s wording. If you quote AI output directly, cite it; see how to cite ChatGPT.
Privacy
Unpublished papers, interview transcripts and data under ethics approval should not go into a consumer chatbot whose terms allow training on your input. Check the tool’s data policy before you upload anything you could not post publicly.