Why We Choose Human Over Machine Translation

Machine translation is good. It is much better than it was five years ago, and it will be better next year than it is today. Yet our company is built almost entirely on human translators. Here is why.

Machine translation is fluent but not accurate

Modern neural machine translation produces text that reads naturally. The trap is that fluent text feels accurate even when it is not. A reader who does not speak the source language has no way to verify whether the meaning has been preserved.

The cost of an error is asymmetric

If a marketing email contains an awkward sentence, the cost is mild embarrassment. If a contract clause is mistranslated, the cost can be a lawsuit. If a medical record is wrong, the cost can be a life. We work in domains where the asymmetry of errors makes machine translation a poor bet.

What we do use machines for

We use translation memory tools, terminology databases, and spell-checkers. We do not use raw machine output as a substitute for translator judgement.

This will likely change as the technology matures. For now, we err on the side of human expertise.