Abstract
This study investigates how large language models (LLMs), specifically GPT-4, can analyze, explain, and enrich concepts in the International Vocabulary of Metrology (VIM). It uses a white-box approach, in which “white-box” qualifies the input, which is fully known and traceable to a validated normative source, while the model itself remains opaque: the material supplied consisted only of VIM definitions and graphical representations of the relationships between VIM terms, together with the relationship types of the informative Annex A of the VIM. The outputs were analyzed qualitatively by structured expert cross-review against three criteria: fidelity to VIM concepts, terminological consistency, and explanatory quality. Across the seven cases examined (four topology diagrams and three VIM concepts), GPT-4 generated structured descriptions and rendered the hierarchical and functional relationships between concepts. Limitations were also observed: occasional non-normative synonyms and purely qualitative descriptions of quantitative relationships. Output variability, although not measured here, is a known LLM property that calls for expert review. Under the conditions tested, GPT-4 appears to be a promising semantic assistant for studying normative vocabularies, with the controlled input context supporting a traceable assessment of its capabilities and gaps. These observations also suggest future directions, which were not tested here, such as AI-assisted teaching, standardization, and terminological interoperability.