Using AI Translation Tools Without Awkward Output
AI translation is fast and mostly correct, but mostly correct is not enough for content that ships. Here is how I get usable translations. I work on content that reaches audiences in three languages and I do not speak all of them. AI translation tools fill the gap, but the gap between mostly correct translation and translation I can actually publish is wider than I expected. Mostly correct translations have word order that feels slightly off, idioms translated literally, and a tone that sounds like a textbook rather than a person. Here is the workflow I use to close that gap. Translate in the Smallest Useful Unit My first instinct was to dump an entire article into the translator and accept the output. That worked for personal reading and failed for anything I shipped. Paragraphs translated in isolation often had references that pointed back to a previous paragraph the model had already moved past. By the end of the article, pronouns and logical connectors were tangled. I now translate in the smallest unit I can act on independently, usually a paragraph or a short section. I keep the source text next to the translation so I can spot drift. For each unit, I ask the model to preserve references and to flag any phrase that requires cultural adaptation rather than literal translation. The output is slower but stays coherent across a long piece.