Failure mode where LLMs produce plausible but incorrect answers. Often caused by knowledge cut-offs (e.g. GPT-4 trained on data through April 2023) or by lacking access to proprietary knowledge. RAG is a common mitigation.