Researchers find that retrieval-augmented AI models still struggle to cite source evidence, with only 22-40% source adherence compared to 0% for non-RAG models.

Study evaluated six LLMs on 90 Stack Overflow questions using three programming textbooks as authoritative sources

Non-RAG models showed 0% median source adherence, while baseline RAG systems achieved only 22-40% depending on the model

Researchers propose illocutionary explanation planning approach to improve faithfulness and traceability in LLM-generated explanations

Focus on making AI explanations scrutable so users can verify claims are actually supported by source evidence