This session traces seventy years of MT history to show that each leap in fluency, from rule-based systems through to today's LLMs, never removed the risk of confident, hard-to-spot errors, which matters most in legal, medical, and government settings. It then defines quality estimation as predicting quality from only the source and MT output, and walks through its evolution from hand-crafted features to learned metrics like COMET and today's LLM-as-judge approaches. The session closes on QE's remaining limits and a translation-literacy message: the future is not translation without humans, but translation with better instruments and humans who know how to read them.