AI for Systematic Review Searching

Systematic reviews require explicit questions, eligibility criteria, comprehensive searches, study selection, data extraction, and synthesis. These steps create a traceable evidence base, so speed cannot come at the cost of unreported omissions. AI can support several tasks: suggesting synonyms, deduplicating records, ranking likely relevant abstracts, screening text, or extracting fields. Each task has a different failure mode. A search assistant can omit a database term; an active-learning tool can rank an eligible article low; a language model can invent an effect estimate. PRISMA 2020 asks authors to report the search sources and dates, full search strategies, and the selection process, including how many reviewers screened records, whether they worked independently, and details of automation tools when used. The guideline is a reporting standard, not a substitute for designing the review. ASReview documentation describes researcher-in-the-loop screening in which people label records and the model prioritizes the next likely relevant record. A BMJ Open methods paper similarly describes an ASReview workflow with human screening. These examples show how AI can prioritize work while preserving human decisions. A review team should define the protocol before using automation, decide which stages are supported, and validate the tool on known included studies or a representative sample. Screening prioritization is not the same as safely excluding records. If the process stops after a chosen number of irrelevant records, estimate the risk of missed eligible studies and report the stopping rule. Full-text eligibility and data extraction may require a different level of review than title/abstract sorting. Keep the search reproducible: preserve database names, dates, exact queries, deduplication rules, software and model versions, prompts, decisions, and reviewer corrections. Use a second human reviewer when the protocol or discipline requires it. Audit false exclusions and disagreement, and update searches before publication when appropriate. AI can help manage volume, but the review authors remain responsible for coverage, accuracy, and transparent reporting.