Large language models (LLMs) are increasingly relying on external knowledge sources to answer complex questions. But are they truly *understanding* this information, or just cleverly mimicking it? New research explores how LLMs choose which external facts to prioritize when faced with conflicting or irrelevant data—drawing a fascinating analogy to legal chains of evidence. Researchers discovered that LLMs perform best when provided with a cohesive set of interconnected facts, much like a lawyer building a case. They call this a “Chain of Evidence” (CoE). The study tested various LLMs, including GPT-3.5, GPT-4, Llama2-13B, Llama3-70B, and Qwen2.5-32B, by feeding them questions along with supporting evidence, some accurate and some not. The results reveal that LLMs are more likely to follow a strong CoE, even if it leads to a factually incorrect answer, suggesting a potential vulnerability to manipulation. Furthermore, LLMs with stronger reasoning capabilities exhibited greater resilience against misleading information, clinging to facts within the CoE despite conflicting data. This research opens intriguing avenues for enhancing retrieval methods in AI, emphasizing the importance of structured, interconnected knowledge. However, it also raises concerns about potential misuse, highlighting the need to ensure the accuracy and trustworthiness of external data fed to LLMs. The balance between leveraging external knowledge and safeguarding against manipulation will be a critical challenge as LLMs continue to evolve.