Translation Did Not Disappear. It Moved Up the Chain and Shed a Tier.

Automated translation crossed a practical threshold when its output became consistently good enough to correct rather than replace. That distinction is the whole story. Editing a flawed draft and producing a translation from scratch are different jobs with different rates and different speeds, and once the draft is reliable enough, the second job largely stops being commissioned for general-purpose content. The volume of translated material has risen sharply, because content that was never worth translating at professional rates now is, and that expansion is real. It has not distributed evenly across the people who used to do the work. The tier that contracted hardest is the one handling routine, high-volume, low-consequence material: product descriptions, support articles, internal documentation, listings. That work sustained a great many freelance careers and functioned as the entry point where translators built experience. Its shift to a post-editing model compressed rates and, more consequentially, removed the rung on which people learned. The tiers that held up are the ones where the cost of an error is high or where the task is not really translation: regulated and legal material, marketing copy that must be re-created rather than converted, and the specialist domains where the client is buying accountability along with words. The demand that grew fastest is for judgement about when the automated output is not acceptable, which is harder to supply than it sounds. Machine translation fails in a particular way — fluently. The output reads naturally and is wrong, which is more dangerous than output that reads badly and is wrong, because a fluent error survives review by anyone who does not know the source language well. Detecting that reliably requires the same expertise that producing the translation required, and it takes longer per unit of confidence than clients generally expect to pay for. For smaller languages the picture is worse than the headline figures imply, and it compounds. Quality tracks the volume of training material available, so languages with fewer speakers get weaker output, and their translators face the same downward rate pressure with a worse draft to correct. Organisations publishing into those markets should be explicit about where automated output is acceptable — reversible, low-stakes, internal — and where it is not, rather than applying one policy to a portfolio in which the tool's reliability varies by an order of magnitude between languages.