Over the past several years, the IAEA’s International Nuclear Information System (INIS) has increasingly relied on artificial intelligence to extract, summarize, and index scientific literature. AI now supports the harvesting of bibliographic metadata from source documents, as well as quality assurance and subject indexing, improving efficiency, speed, and consistency. Recently, INIS was provided with 115 historical works of the Austrian Physicist Erwin Schrödinger, for inclusion in its repository. This offered an opportunity to test the limits of AI metadata extraction. Using a state-of-the-art model, a very large, restrictive prompt, and human-in-the-loop review, the approach sought to eliminate uncertainty. However, the experiment revealed a paradox. Instead of eliminating uncertainty, the process exposed its structural persistence. This presentation proposes a “Metadata Uncertainty Principle” to help understand the limits of AI extracted metadata and for designing systems to govern it.