When faced with the task of translating a very long document, using AI to support this time-consuming work is an obvious option. However, simply asking an AI system to translate a text from one language into another, without giving it any further instructions, can lead to errors and generally unsatisfactory results. A better approach is to use AI as a translation partner while continuing to guide the process through human decisions and quality checks.
We first gave the AI clear instructions for the translation: the source and target languages, the need to preserve the author’s first-person perspective, and the requirement to retain repetitions and stylistic characteristics. We also specified that the translation should remain as faithful as possible to the original and should not simplify, shorten or independently “improve” the text.
Consistent terminology throughout the document was equally important. This applied to specialist terms as well as terminology from education, law and business. Once a term had been chosen, it should be used consistently wherever possible. Where no direct equivalent existed in the target language, a decision had to be made as to whether the original term should be retained, explained or replaced with the closest available translation.
Translating the entire document in a single step did not appear either efficient or sensible. With a very long text, the risk of inconsistencies, omitted nuances or inaccurate wording increases. We therefore worked through the text section by section. This made it possible to review the quality of the translation as we went along and to apply earlier language choices consistently in later chapters.
Once the first translation was complete, we read through the translated text in full ourselves and marked passages that sounded unnatural, overly literal or potentially misleading in a cultural context. We then used the AI to discuss possible alternatives and explain the semantic differences between different formulations. This was particularly useful for ambiguous expressions, idioms, historical terms and concepts that had no direct equivalent in the target language.
After the full translation was finished, the AI carried out a systematic bilingual review. The aim was not only to identify clear translation errors, but above all to flag so-called “grey areas”: passages where several translations might be possible, or where meaning, tone, cultural context or terminology required closer human review.
The AI then produced a separate document containing the flagged passages, which we could work through point by point. This made the final quality-control stage much more targeted. Instead of having to review the entire text sentence by sentence again, we could focus first on the passages that were particularly sensitive from a linguistic or contextual perspective.
AI was therefore not used as a “one-click” translation machine, but as a translator, terminology adviser and bilingual review partner. The final linguistic and editorial decisions, however, remained with the human editor.
By Andrew Colclough