Large Language Models in der systematischen Literaturrecherche – eine Evidenzübersicht

Abstract

Background: The methodology of systematic literature searching requires that the information retrieval process has a high recall and is as transparent and reproducible as possible. The introduction of large language models (LLMs) such as ChatGPT raised the expectations for automation of evidence synthesis processes through artificial intelligence (AI). However, assessments of the usefulness of LLMs for systematic searching are heterogeneous. This narrative review examines the current evidence on the use of LLM tools compared to systematic searches performed by humans (as of August 2025). Results: The majority of studies focus on two areas of application: the creation of Boolean search strategies by LLMs and the generation of comprehensive literature lists using AI-supported search platforms (Elicit, Consensus). In both cases, AI tools achieved insufficient recall rates compared to traditional systematic search methods. However, AI-supported search platforms were able to identify additional studies that were not found by Boolean search strategies. Few studies investigated the use of LLMs for error detection in database search strategies. AI was able to find errors, but there were problems in creating improved search strategies. Conclusion: Based on the available evidence, AI-supported methods should at most be used to complement established methods of systematic literature research. On their own, they neither achieve the necessary high recall nor are their results reproducible. However, there are also significant gaps in the evidence. Independent evaluations and critical assessment of AI tools by users remain essential.

Ähnliche Arbeiten