Explanations, Evidence, and Overreliance in Retrieval-Augmented Generation: A Critical Literature Review

The presence of evidence in output generated by retrieval-augmented generation (RAG) does not guarantee that the information is correct or that the users will evaluate it properly. This research examines the ethical relationship between explanations, source attribution, and overreliance in artificial intelligence systems. It synthesizes scholarly literature of retrieval and attribution, explanation faithfulness, and human-artificial intelligence interaction. It highlights four properties: factual correctness, evidential support, source credibility, and explanatory faithfulness. The study indicates evidence that explanations can increase acceptance of incorrect answers, however useful sources and lower verification costs can support appropriate reliance. The findings observed in this research do not establish a universal standard of citations or explanations in deployed RAG systems. The review argues that users should maintain the practical ability to question recommendations in responsibly designed systems. A conceptual framework is proposed by connecting evidence presentation to verification effort, opportunities for contestation, and responsibility. Key issues include identifying unsupported claims and source disagreement, distinguishing post-hoc attribution from causal explanation, and evaluating risk controls based on decision outcomes as well as reported trust. The contribution is an interpretation and research agenda, further experiments and empirical research needs to be done to extend this research.