Information retrieval reliability in large language models: a study of source verification

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Large language models (LLMs) have come into widespread use in recent years across domains ranging from education and journalism to academic research and everyday information seeking. Their ability to produce fluent, coherent-sounding answers in natural language creates the impression that these answers are accurate and trustworthy, yet the literature offers strong evidence that this impression does not always hold: models occasionally produce content that is factually wrong or unsupported by any source, a phenomenon known as hallucination. Drawing on recent academic work, this paper reviews reliability problems in the information retrieval processes of large language models and the source verification approaches developed to mitigate them. The paper examines, in turn, the concept of hallucination; the retrieval-augmented generation (RAG) architecture; methods for estimating source reliability; self-verification techniques such as chain-of-verification; semantic uncertainty estimation; context–memory conflict problems; and benchmark studies aimed at hallucination detection. It also discusses alternative approaches that propose generating information from the model itself rather than retrieving it from outside, considering how retrieval-based and generation-based strategies might complement one another. The findings indicate that while the RAG architecture reduces hallucination substantially, it can introduce new reliability problems tied to the quality, currency, and diversity of the sources used; that explicitly modelling source reliability, multi-step verification mechanisms, and independent benchmark datasets are effective in mitigating these problems. Written from the perspective of a high-school researcher, the paper concludes with criteria that students and general users should attend to when evaluating AI-assisted information sources, with particular attention to the difficulties faced in lower-resource languages and with recommendations for educational settings.