Google NotebookLM Audio Summaries Contained Fabricated Quotes and Data

Google's NotebookLM, launched as an AI-powered research assistant, offers a unique feature that converts uploaded documents into podcast-style audio summaries with two AI hosts discussing the content. In early 2025, researchers and users began documenting instances where these audio summaries contained fabricated information not present in the source documents. The fabrications took multiple forms, including quotes attributed to real individuals that were never made, statistical claims not found in the source material, and contextual details invented to enhance the conversational flow of the audio content. The AI-generated hosts would cite these fabricated elements with the same confidence and natural speech patterns as legitimate information from the documents. Researchers noted particular concern about the format of these hallucinations. Unlike text-based AI outputs where users might more readily question unusual claims, the conversational audio format with natural speech patterns and discussion between two AI hosts created an enhanced sense of credibility. Users reported finding the fabricated content convincing, especially when the AI hosts would reference the invented quotes or statistics multiple times throughout a summary. Google had marketed NotebookLM's audio feature as a way to make complex documents more accessible and engaging, with promotional materials emphasizing the tool's ability to accurately summarize and discuss uploaded content. The company's documentation suggested users could rely on the summaries for research and learning purposes, though it did include general disclaimers about AI-generated content accuracy. The incidents highlighted broader challenges with large language model hallucinations in audio formats, where the natural conversational style may increase user trust compared to traditional text outputs. Users across academic, professional, and personal use cases reported discovering the fabrications only when they later fact-checked claims against their original documents, often after sharing or acting on the incorrect information.