AI tools in evidence synthesis - Searching for Systematic Reviews & Evidence Synthesis - LibGuides at King’s College London

Research demonstrates that ChatGPT can effectively support multiple stages of the systematic review process, including Boolean query generation, abstract screening, full-text data extraction, and thematic content analysis.

• Search Strategy: It aids in preparing Boolean research terms and collecting articles by refining search queries.
• Screening: It assists in categorizing abstracts and filtering full-text articles based on inclusion criteria.
• Data Extraction: It is used to extract specific information from randomized controlled trials and conduct Risk of Bias analyses.
• Analysis: It facilitates content analysis to identify trends, challenges, and gaps in the evidence.

While these applications show high accuracy in filtering and classification, tools remain insufficiently reliable to completely replace human judgment and must be used as adjuncts with transparent reporting.

This King’s College London LibGuide provides a practical orientation to using generative AI, particularly ChatGPT, within systematic review and evidence synthesis workflows. It maps LLM-assisted tasks across the review pipeline, including Boolean query generation, abstract screening, full-text data extraction, and thematic analysis. The central framing is an augmented, human-in-the-loop model: the language model can help with repetitive pattern recognition, drafting, and code suggestion, while reviewers retain responsibility for query logic, eligibility decisions, risk-of-bias appraisal, and interpretive judgment.

A key insight is that the usefulness of AI in evidence synthesis is highly workflow-specific and depends on careful prompt design, task scoping, and verification. For search construction, ChatGPT can help translate research questions into Boolean strings, but outputs still need checking against database-specific syntax, field tags, and subject-matter coverage. For screening and extraction, the guide implicitly highlights the need for explicit criteria, structured outputs, and audit trails to reduce missed records or hallucinated details. For thematic analysis, it suggests that LLMs can accelerate familiarization and coding, but themes must remain grounded in reviewer-defined analytic frameworks and validated against the source texts.

The material matters because systematic reviews face growing pressure to scale with expanding evidence bases without sacrificing methodological rigor. It is useful for practitioners seeking concrete AI use cases and for methodologists thinking about how to document LLM assistance transparently. Its broader contribution is to position AI tools as productivity and consistency aids rather than epistemic arbiters: responsible adoption depends on prompt versioning, model disclosure, error checking, data governance, and clear reporting of where AI output was used and how it was validated.