So we do at least agree on the fact that quality of human language <-> embeddings transition depends on how good target language is represented in the training

So we do at least agree on the fact that quality of human language <-> embeddings transition depends on how good target language is represented in the training dataset? Even if it happens at the edge, on every conversation turn, I may assume non captured small subtleties of meaning over time can accumulate into significant error.