Skip to content
A global perspective on digital financeAbout The Token Press
CRYPTO. FINANCE. WHAT’S NEXT.The Token PressYour daily briefing ↗

Beginner · 6 min read

How to Check an AI Translation Before Publishing

Check names, numbers, missing instructions and language choices in an AI translation, with an original workshop exercise and a review prompt.

The Token Press education desk · Updated

Learning path: Build with AI: ChatGPT, Astra, Gemini and a Practical Learning Path →

What can you check if you do not speak the language?

You can check whether names, amounts, dates, links and required instructions survived an AI translation. You cannot establish natural wording or correct meaning in an unfamiliar language simply because the result looks polished. Treat a translation as a draft until the checks appropriate to its use are complete.

This guide gives you a practical review process for a website, app or short learning handout. It is different from checking whether an AI answer cites a real source: a translation can contain every expected number yet reverse an instruction. For material that readers will act on, arrange review by someone who understands both languages and the subject. Do not label an automated comparison as human review.

The exercise below uses an invented workshop notice. It requires no real customer data or paid translation account, and does not claim that any particular model produced the deliberately flawed version.

Prepare the source, audience and terminology

Start with the approved source text and a version number. Resolve unclear wording before translation: does 'deposit' mean a refundable reservation payment or a crypto transfer? Does 'by Friday' include Friday, and in which timezone? If the source is ambiguous, asking the model to sound more confident will not settle the intended meaning.

Specify the reader, language variety and tone. A public workshop notice may need plain, welcoming language; a button needs a short action label that matches the interface. W3C explains that language tags can distinguish useful regional variants, such as es-ES and es-419. Choose the language actually used by the audience rather than assuming that one country implies one language.

Make a small terminology list: approved product names, terms with a particular meaning, and labels that must match the app. If you do not know the target-language equivalents, leave them for a competent reviewer instead of inventing a glossary. Google Cloud supports glossaries for preferred term translations, but its documentation also identifies matching rules and exceptions. Supplying a glossary is a control to test, not proof that every occurrence was handled correctly.

Use public or fictional text for a first trial. Before uploading private material, check the chosen service and account's data terms; remove customer identifiers that the task does not need.

Ask for a traceable translation draft

Give each paragraph or interface label an ID, then ask for the same IDs in the output. Keep the source beside the translation so that a reviewer can locate an omission without searching a long document. Ask the model to report uncertainties separately instead of silently adding facts or resolving unclear policy.

The following prompt is a reusable brief, not a claim that a test has been run. Replace the bracketed fields with approved information and attach only the text you intend to translate.

Prompt to adapt: Translate the approved source below for [audience and target language/region]. Preserve each segment ID. Keep names, currency, quantities, URLs, placeholders and the meaning of every condition or negative instruction. Use the approved terminology list where applicable. Do not add facts or convert currencies. Return: (1) aligned source and target segments, (2) a list of ambiguous source phrases, and (3) a checklist of facts to verify. Do not claim the output has been reviewed by a human. Source version: [version]. Source and terminology: [text].

Use two separate review passes

First compare coverage and fixed facts: every source segment appears once; names and links match; quantities retain their units; dates are unambiguous; and placeholders such as {booking_id} remain intact. The visible formatting of a number may legitimately change by locale, so compare its value and meaning, not only its characters.

Then check meaning and usability with a bilingual reviewer: who must do what, whether an action is allowed or prohibited, which conditions apply, and whether the wording suits the audience. Ask the reviewer to flag specific source/target segments and proposed corrections. DeepL's evaluation documentation separates accuracy, fluency and style, which helps explain why grammatical text can still convey the wrong instruction.

Some tools offer automated quality reports. DeepL documents a file-translation evaluation option with plan, language and platform restrictions, including beta availability in some delivery routes. It flags potential issues rather than automatically correcting them. Check the current support page for your account instead of assuming the feature is included.

A second model or a back-translation can raise useful questions, but do not treat agreement as independent proof: both outputs can miss the same problem. Our recommended acceptance rule is to resolve each material discrepancy against the approved source and record who actually performed the language review.

Five-minute practice: catch four changed facts

Approved fictional source: 'Workshop on 18 December 2026 at 18:30 Dubai time (UTC+04:00). Maximum 24 attendees. Price: AED 125 per person. Do not pay until your place is confirmed by email. Bring a laptop; chargers are not provided.'

For a language-independent checking exercise, imagine a target-language draft with this deliberately flawed English meaning summary: 'Workshop on 18 December 2026 at 18:30 UTC. Maximum 24 attendees. Price: USD 125 per person. Pay now to confirm your place. Bring a laptop.' This is an invented error sample, not a translation benchmark or an actual back-translation.

List the differences before reading the answer. Four material checks fail: the timezone changed; AED became USD; the payment instruction changed from waiting for confirmation to paying immediately; and the charger warning disappeared. The date and capacity matching do not make the draft acceptable.

Check the arithmetic independently: 24 places at AED 125 would total AED 3,000 if all places were sold at that price, before any other charges. This does not authorize a payment. The original 18:30 time at UTC+04:00 corresponds to 14:30 UTC, so preserving 18:30 while relabeling it UTC moves the event by four hours.

Now change only the approved capacity from 24 to 20. The revised maximum gross amount is AED 2,500; the date, local start time, currency and payment condition stay the same. Have the reviewer check the changed segment and any dependent text rather than accepting an unexamined rewrite of the whole notice.

Visual explainerThe digits can match while the meaning changes
Time18:30 DubaiUTC+04:00 means 14:30 UTC, not 18:30 UTC.
PriceAED 125Keeping 125 while changing the currency changes the offer.
ActionWait to payPaying now reverses the approved confirmation condition.
Also restore the missing charger warning before approving the draft.Original fictional workshop exercise. No model accuracy or human-review claim.

Check the delivered page and keep the review record

Read the translation in its final layout. Look for clipped buttons, hidden conditions, broken links and a reading order that changes on mobile. Confirm that language and text-direction settings match the content. A sentence approved in a document can still become unusable when its final line is hidden inside a fixed-height card.

Save the approved source version, translation version, unresolved questions and review outcome together. If the source changes later, identify the affected target segments. 'AI-assisted draft, language review pending' and 'reviewed by [actual reviewer] on [actual date]' describe different states; only use a review claim supported by work that happened.

For medical, legal or other consequential instructions, use a qualified translation/review process appropriate to that subject. Passing this short exercise does not certify the accuracy of a real translation.

  • All source segments and required conditions are present.
  • Names, numbers, currency, timezones and links match the approved meaning.
  • Uncertainties and terminology choices have been resolved by a suitable reviewer.
  • The final mobile page or file remains readable and functional.
  • The recorded review describes actual checks, not an invented endorsement.

Sources & further reading

Educational content. Examples are illustrative. Protocols, products and rules can change; consult the linked documentation for current details.

Explore the newsroom

Connect what you have learned with the latest reporting.