How to Write a Literature Review with AI

A reliable workflow has five stages. First, define a focused question and inclusion criteria, such as years, populations and study types. Health researchers often use the PICO frame (population, intervention, comparison, outcome). Second, search: combine standard databases (Google Scholar, PubMed, Scopus, Web of Science or your field's database) with AI research tools such as Elicit, Consensus or Semantic Scholar, which search indexes of real papers, and use citation chaining to follow references backward and forward. Third, screen titles and abstracts against your criteria, and record why you excluded papers. Fourth, extract: build a matrix with one row per study and columns such as authors and year, question, method, sample, key findings and limitations. AI can draft entries from full text you provide, but check each cell against the paper. Fifth, synthesize. A literature review is not a list of summaries; it groups studies by theme, agreement, method or gap, and explains what the field knows and does not know. AI can suggest groupings from your matrix; deciding which ones are meaningful is your job. The critical step is citation checking. General chatbots without search can invent references with realistic authors, titles and journals. The risk is real outside academia too: in the 2023 US case Mata v. Avianca, lawyers were sanctioned for filing a brief with court cases ChatGPT had invented. For each reference, confirm the DOI resolves, the title, authors, year and journal match, and the paper actually says what you claim. A common misconception is that tools built on real databases cannot be wrong. They avoid invented papers but can still misstate findings. Also check publisher terms before uploading PDFs, follow reporting standards such as PRISMA for systematic reviews, and disclose AI use wherever your journal or institution requires it.