The point here is that if machine translation becomes so good that you become so confident about it that you don't think you need to cross-check or double-check the automated translation, then this can lead to wrong and potentially dangerous actions/decisions [the assumption here is that merely "almost-flawless" automated language translation can create this high level of confidence or trust in a human user, and that "fully-flawless" level of accuracy isn't required]. Since the translation service is not 100% accurate but only almost-100%, it will inevitably make occasional errors/mistakes. But since the human user has complete confidence or faith in this service and so he doesn't feel the need to get the automated translation checked [manually - by a human translator], the error/mistake can be [silently] accepted by the human user as if it were the correct translation [that is, the human won't even realize that there's any flaw in the translation]. This can lead to erroneous actions or decisions. If doctors rely solely on such "almost-perfect" computer translation, serious medical blunders can occur.
An example of very good automated translation is below. How can I claim that the translation correctly depicts what was originally written in Ukrainian? I read a few news stories [in English] about this Ukrainian-language webpage. But I cannot be sure that those news outlets got this page translated by a human, or they themselves too relied on Google Translate.
An example of very good automated translation is below. How can I claim that the translation correctly depicts what was originally written in Ukrainian? I read a few news stories [in English] about this Ukrainian-language webpage. But I cannot be sure that those news outlets got this page translated by a human, or they themselves too relied on Google Translate.
Update [7-Oct-18]: Another [representative] scenario where an almost-accurate system can lead to catastrophic outcomes is an imaginary crying-infant detector device which continuously listens to incoming sounds, and can identify sounds that resemble a crying infant. It alerts the parents - who are sitting at some distance - when it detects this crying. Support it's "so" accurate that parents start to depend blindly on it. Suppose its real accuracy is 99.5%, whereas the parents consciously or subconsciously start to assume - based on their real-life experience with the device - that it's "virtually" 100% accurate. Now this can lead to fatal mistakes. What if a particular type of low-volume, intermittent crying by an infant falls in that 0.5% category, and the device doesn't raise an alarm, and the parents falsely assume that all is well, and the infant keeps crying for a long, long time? It's a scary scenario. In any such high-stakes scenarios, any machine-based system better be at least as accurate as a human. Nearly-100% can produce fatal outcomes due to blind faith and by the system's human users. As it's sometimes said, good enough is not good enough.