Back to Blog
    AI Writing November 05, 2026 5 min read

    How to Review an AI Translation Before Publishing

    Review an AI translation in two passes: compare the meaning against the source, then check the translated text in its final format.

    Golden morning mist between layers of rolling hills

    Review an AI translation in two passes. First, put it beside the approved source and check what the reader is told to do, under which conditions, and with which names, numbers, or limits. Then read the translated version on its own, in the actual email, page, or screen where it will appear. A fluent sentence can reverse a condition; an accurate sentence can still break when a placeholder is filled in.

    Start with the version you can approve

    Keep the source revision, the target language and locale, and the intended place of publication together. If the source is still changing, assign a version or date to the text you reviewed. Otherwise a correction in the source can leave a previously checked translation out of date.

    Decide who can approve meaning in both languages. A person who reads only the target language can give valuable feedback on clarity, but they cannot confirm that the translation preserves what the original says. The Bern University of Applied Sciences checklist for machine translation distinguishes suitability for the audience, terminology, and when to involve a professional translator.

    Pass one: compare the meaning across languages

    Use a row for each meaningful unit: a sentence, a warning, a button and its accompanying text, or two sentences that share an exception. Read the source before the translation. Ask what a reader could reasonably conclude from each, rather than trying to match individual words.

    Test the rule with two readers

    For any condition, imagine one reader who meets it and one who does not. Would they take the same action after reading the source and the translation? This catches problems that a word checklist can miss. A translation might contain an equivalent of “not” and still apply it to the wrong action. Research on negation in machine translation distinguishes omission, reversal, and changes to what is negated.

    Check what was added or lost

    Mark amounts, dates, units, named entities, exceptions, and the force of verbs such as “must,” “may,” and “can.” Follow references to other sentences and labels. Something can move to the next sentence without being lost, but it must still be available where the reader needs it. The MQM community error taxonomy separates additions, omissions, and mistranslations from fluency issues. For the related single-language task, use the guide to preserving meaning during a rewrite.

    A worked example: a pickup reminder

    This is an invented bilingual example, not an observed model output. The English source reads:

    Pick up your order by 6 p.m. on 12 May. Do not come until you receive the “Ready for pickup” message. Bring order code {order_code}.

    A plausible but defective Spanish draft reads:

    Recoge tu pedido antes de las 6 p. m. del 12 de mayo. Ven aunque no hayas recibido el mensaje «Listo para recoger». Trae el código {codigo_pedido}.

    Review it by reader consequence:

    ItemWhat went wrongWhat to do
    Pickup window“Antes de” can imply strictly before 6 p.m.; the source says by 6 p.m.Preserve the deadline with “a más tardar a las 6 p. m.”
    Ready message“Aunque no” tells people to come even without the message.Restore the wait condition. Confirm the message's exact label in the Spanish interface.
    Order codeThe placeholder name changed. If the template expects the source token, the value may not appear.Keep the literal token unless the template owner confirms a different target token. Preview it with a safe sample value.

    If the interface owner confirms the Spanish label, an edited version could be:

    Recoge tu pedido a más tardar a las 6 p. m. del 12 de mayo. No vengas hasta recibir el mensaje «Listo para recoger». Trae el código {order_code}.

    An editor can now test two scenarios: a customer with the message can come before the deadline; a customer without it must wait. If the deadline or time zone is unclear in the original, ask the source owner rather than solving the ambiguity in only one language.

    Pass two: review the target in its final format

    Read the target as a standalone piece of writing. Check whether the term readers see in the text matches what they will find in the app, help page, or form. A translation glossary can record approved labels across languages; an editorial style sheet for AI-assisted writing can keep recurring choices consistent after they are approved.

    Check locale without changing the underlying facts

    Use the destination's conventions for dates and decimal separators where appropriate. Keep currency and units explicit. A formatting change is different from converting an amount or moving a deadline into a different time zone. If conversion is intended, record the rule and get it checked against the source.

    Preview what the reader will see

    Open the actual target page or message. Fill tokens with representative values. Follow links, inspect line wrapping on a narrow screen, and check whether footnotes, warnings, or labels are still visible. The Philadelphia translation quality handbook gives examples of why fonts, symbols, and layout deserve their own review. Reading a text export alone cannot confirm the final layout.

    Keep a short record of release decisions

    For each unresolved issue, record the source revision, target passage, reader impact, suggested correction, and who must confirm it. Block release when a condition is reversed, a required step disappears, a link goes to the wrong destination, or a reviewer cannot establish what the source means. Set aside optional wording preferences after the substantive problems are resolved. For material where errors could cause serious harm, use the applicable review process and qualified language and subject expertise.

    FAQs

    Can back translation replace bilingual review?

    No. It can raise a question, but it adds another interpretation. Resolve the question by comparing the actual target wording with the source and its context.

    Do I need to fix every phrase that sounds different from the source?

    No. A natural target sentence may use different word order or split a long sentence. Check whether the same reader can reach the same conclusion and take the same action.

    Humanizer.help Editorial Team

    About Humanizer.help Editorial Team

    The Humanizer.help editorial team writes practical AI writing guides for students, ESL writers, and academic drafts that need clearer, more natural English.

    Related Articles