Briva-Iglesias and Ferre-Fernandez show that multilingual access to cultural heritage depends on terminological governance rather than generic linguistic fluency. Their comparison of neural machine translation, simply prompted language models and glossary-augmented retrieval demonstrates that a modest domain glossary can substantially improve exact terminology while preserving overall textual quality. The methodological lesson is economical and important: institutional AI does not always require model retraining or elaborate infrastructure; carefully curated lexical resources and targeted human evaluation can produce decisive gains. Rock-art discourse provides a stringent test because small variations in motif labels or chronocultural categories can distort interpretation and propagate error through education, indexing and public communication. Translation emerges as a knowledge-representation problem in which lexical consistency secures the mobility of specialised concepts across languages. The bridge to conceptual engineering is direct: a term is not merely translated but maintained across contexts, and its successful circulation depends on an infrastructure that couples machine generation with explicit, revisable semantic constraints.
