Practical guide
Check an AI translation before anyone acts on its meaning
An AI translation can look fluent while changing a condition, a number, a warning or who must do something. Verification is therefore a meaning check against the source, by a person competent in the relevant language and context, before a consequential translated passage is used. If that reviewer is unavailable, label the translation as unverified and keep it out of the decision path. A.
An AI translation can look fluent while changing a condition, a number, a warning or who must do something. Verification is therefore a meaning check against the source, by a person competent in the relevant language and context, before a consequential translated passage is used. If that reviewer is unavailable, label the translation as unverified and keep it out of the decision path. A back-translation or a second model output can expose questions, but neither replaces a competent comparison.
Define what the translated text will be used for
Imagine a synthetic project team receiving a Spanish-language supplier update. A worker asks an AI assistant for an English summary for a manager. The source says a delivery date is provisional pending inspection; the English draft says delivery is confirmed. The sentence sounds natural, and a spell checker will not notice the changed obligation. Before reviewing, the worker marks the intended use: is this merely a rough reading aid, or will the manager use it to commit a customer date? The higher-consequence use requires an approved translation or a qualified reviewer who checks the source, the audience and the relevant terminology.
The European Commission's explanation of machine translation on Europa says accuracy can vary between texts and language pairs, and distinguishes automatically translated text from authentic language versions. Its AI translation and language tools guidance describes machine output as a starting point for further work, not a substitute for a quality-controlled final version. These are examples of the Commission's own publishing practice, not a rule that every organization must adopt.
Sources: [1] [2]
Compare high-consequence segments side by side
The reviewer makes a small bilingual check sheet, not a generic “translation quality: good” score. One row contains the source segment, the proposed English segment, why the meaning matters, the authoritative terminology if one exists, the corrected text, and who approved it. In the synthetic supplier update, row one covers the provisional date; row two covers whether inspection is a condition or an already completed event; row three checks the product identifier and quantity. The final summary becomes: “The supplier proposes a date, subject to inspection; do not confirm the customer commitment yet.” The exact Spanish wording and English rendering are fictional teaching material, not a claim about a real supplier.
W3C's guidance for reviewing WAI translations calls for comparing the translation with the original, preserving meaning and checking terminology appropriate to the target language and region. It also notes that back-translation can sometimes help identify issues while being inaccurate itself. Applied to this work example, the qualified reviewer checks negation, conditional verbs, units, named parties and any locally important terms against the source. If a segment cannot be resolved, the check sheet records the uncertainty and routes it to the document owner instead of smoothing it into a confident sentence.
Sources: [3]
Verify the decision artifact, not only the sentence
After the bilingual comparison, inspect every place the translated claim appears: email draft, slide headline, spreadsheet note and manager summary. A corrected paragraph does little good if the slide still says the date is confirmed. Link the reviewed version of the source and translation, and identify whether later source revisions invalidate the check. For a time-sensitive decision, ask the owner what should happen if the qualified reviewer is unavailable. The honest answer may be a delayed commitment or a summary that explicitly states the unresolved condition.
The principal failure is to infer trustworthiness from fluent prose, a model's self-rating or a round-trip translation alone. The method should also avoid overcorrecting: not every stylistic difference changes a decision, and reviewers can disagree about phrasing while agreeing about the operative condition. Record the material discrepancy and its resolution, rather than presenting a raw model score as proof of equivalence. The W3C process is a useful comparison practice in accessibility publishing, not certification of a specific supplier translation.
This guide concerns human review of AI-assisted language work, not legal translation advice, regulatory compliance or multilingual publication by Two Prune. In a Two Prune assessment, a synthetic bilingual source can create a task-bounded opportunity to observe whether a participant recognizes and routes an unresolved meaning change. The resulting assessment evidence is for human interpretation, not a statement that the participant is professionally accredited as a translator.
Sources
- 1.European Commission's explanation of machine translation on Europa · commission.europa.eu
- 2.AI translation and language tools guidance · translation.ec.europa.eu
- 3.W3C's guidance for reviewing WAI translations · World Wide Web Consortium