I expect the opposite. I would expect an ML/embedding approach to find lots of false positives, because lots of words have close embeddings but are not synonyms

I expect the opposite. I would expect an ML/embedding approach to find lots of false positives, because lots of words have close embeddings but are not synonyms. A strict thesaurus lookup should produce fewer matches. As for ranking the good ones, an embedder might help with that, but then we need a definition of "good". I would argue the most conceptually "unrelated" matches are the "best" ones, so yes, an embedder could quickly determine the farthest vector distance.