Sure, as long as one has some means of predicting the probability of false positives as well as false negatives. Until then, colour me unconvinced of the (f)utility of this approach :)

LLMs do a pretty darn good job of translating other languages, even preserving inflection and tone and rhyme in some cases. Same when translating programming languages. If the training pool is large enough, they should be quite good at it.

No problem marking them as machine translated and keeping track of which have been spot-checked by experts either.

I think "if the training pool is large enough" is a real issue here. We're not talking about living languages with large, properly attested and annotated corpuses.

Indeed, one of the thing you'd probably like the translators to do is identify rare or unique words that can be added to our existing knowledge of these languages.

> I think "if the training pool is large enough" is a real issue here.

It would be really neat to set up something like a wiki populated with the existing translations and machine translations done via LLM, and to periodically re-train the LLM on all the newly manually verified translations and automatically re-run the machine translations after. The whole thing could move incrementally toward high quality output.

This is the same thing as trusting its translation though. Or worse, since there's more "interpretation" happening (as it's basically a summary rather than a whole text, and you can't trust either of those steps).

It's like using a classifier to sort biopsies from most-likely to least-likely to be Really Bad News. A human will get to them all eventually, but you can at least fast-track the ones that are almost certainly Really Bad News.

(I think I just said "triage" in a lot more words, to be honest).