Machine translation improved more between 2015 and 2020 than in the preceding fifty years, and the improvement changed what the errors look like rather than eliminating them. Statistical systems produced output that was obviously broken; neural systems produce fluent sentences that are occasionally, and undetectably, wrong. A negation dropped or a name swapped now arrives inside a paragraph that reads perfectly. For a user who cannot read the source, fluency is the only available signal of quality, and it has become uninformative. The practical consequence is that the technology is safest where a mistake is cheap and least trustworthy where it matters. Systems now warn about their limitations in the abstract while giving no signal about any particular sentence.