Translation AI: Localise Without Losing Meaning

Translation AI can produce a useful first draft, align repeated terminology and shorten the path from an approved source to a multilingual release. It can also reverse a negation, change a legal promise, corrupt a placeholder or confidently choose the wrong meaning of an ordinary word.

The operating pattern is prepare the source → classify the risk → generate → review meaning → test the experience → approve → monitor. Translation is not finished when every string has text. It is finished when a target-language user can complete the same task with the same material information and a safe route to help.

The translation sources and cross-border rules below were checked on 31 July 2026. UK data-protection duties apply across the UK, but equality law differs: the Equality Act 2010 covers Great Britain, while Northern Ireland retains separate anti-discrimination legislation. Welsh Language Standards bind organisations through their applicable compliance notices rather than every UK business automatically. A UK exporter must also check consumer, privacy, language and sector rules in each destination market; “UK compliant” is not a global localisation rule.

Define the Unit of Meaning

Do not send an entire website to a model and call the output localised. Build a content inventory with an owner, purpose, audience, source version, target locale, expiry date and consequence of error.

Separate language from locale. en-GB, fr-FR and fr-CA carry different spelling, dates, addresses, tax language and consumer expectations. Record the writing system and direction too. Arabic or Persian localisation is not complete if the text is translated but the interface still lays out controls left-to-right.

Start by simplifying the approved source. Resolve ambiguous pronouns, inconsistent product names, unexplained abbreviations and sentences containing several conditions. Translation cannot preserve a meaning the source team has not settled.

Build a Controlled Language Asset Layer

Maintain four versioned assets:

• a termbase containing approved source and target terms, definition, prohibited variants, domain, owner and review date;
• a style guide covering voice, formality, inclusive language, punctuation, numbers, dates, units and transliteration;
• a translation memory of approved segments with source and target versions and provenance;
• a do-not-translate list for product names, code, identifiers, legal references and placeholders.

Treat these as governed product data. A translation memory can preserve a previous mistake just as efficiently as a correct phrase. Expire entries when a product, law or policy changes. Do not let the model overwrite an approved term silently.

Protect variables and markup before generation. Test names, plural rules, gender or grammatical agreement, HTML, Markdown, ICU message syntax, links, currency, decimal separators and line breaks. A syntactically valid string can still send money to the wrong amount or tell a user that an action succeeded when it failed.

Route Work by Consequence, Not Volume

Use a policy engine that chooses a workflow, not a model that decides its own authority. Inputs should include content class, locale, personal-data status, freshness, target channel and required reviewer.

For low-risk material, AI may draft and a bilingual editor can sample. For transactional content, review every changed segment and test the full journey. For rights, money, health and safety, require a suitably qualified human to compare source and target in context; independent verification may be appropriate where a single error could cause serious harm.

NHS England's community-language translation and interpreting framework is specific to NHS services in England, but its emphasis on consistent, high-quality provision is a useful reminder that access to essential services cannot be reduced to raw model fluency. AI does not turn a relative, child or unqualified bilingual colleague into an interpreter.

For live speech, show when audio is being captured, which languages are supported and when a person is joining. Preserve turn-taking and speaker identity. Never hide low confidence behind a smooth synthetic voice. Provide a trained interpreter route for consent, diagnosis, legal advice, safeguarding, complaints and other consequential exchanges.

Evaluate Meaning, Not Just Similarity

Automated similarity scores can help compare experiments, but they do not establish that a release is correct. Build a bilingual test set from real tasks, including short labels, long-form content, placeholders, ambiguous terms, negative statements, numbers and culturally sensitive material.

Human reviewers should classify errors consistently:

• addition, omission or changed meaning;
• incorrect terminology or named entity;
• number, unit, date, currency or polarity error;
• grammar, fluency, register or tone problem;
• broken placeholder, link, markup or layout;
• discriminatory, offensive or culturally inappropriate wording;
• content that is accurate linguistically but wrong for the target rule or process.

Weight severity by consequence. One mistranslated allergy warning matters more than several awkward marketing sentences. Report critical errors separately rather than averaging them away.

Use blind comparison where practical and record reviewer language pair, domain competence and decision. Measure first-pass acceptance, critical and major errors per reviewed words, approved-term adherence, edit distance, reviewer time, reopened defects and user task completion. Break results down by locale, content class, engine/version and channel.

The ICO distinguishes data accuracy from model statistical accuracy. Its AI accuracy guidance explains that AI outputs involving personal data still engage fairness and accuracy considerations. Its broader accuracy principle requires reasonable steps to keep factual personal information accurate and to handle challenges. A translated allegation, opinion or case note must preserve its status and source; it must not become an asserted fact.

Test the Localised Product

Reviewers need the rendered experience, not only a spreadsheet. Test at realistic mobile and desktop widths with the target font and production data.

Check:

• text expansion, truncation, wrapping and overlapping controls;
• right-to-left order, mirrored navigation and mixed-direction numbers;
• input methods, search, sorting and address formats;
• plural, gender and grammatical cases at boundary values;
• currency, dates, time zones, telephone numbers and measurements;
• translated links, attachments, emails, PDFs and error messages;
• screen-reader pronunciation and language switching;
• subtitles, transcripts and audio timing;
• fallback when a translation is missing or stale.

W3C's WCAG 2.2 Recommendation includes requirements for identifying the language of a page and changes of language within content, alongside captions and other accessibility criteria. W3C's HTML language guidance explains how to declare processing language. Correct lang metadata helps assistive technology choose pronunciation; it does not repair a poor translation.

Government guidance on accessible communication formats also distinguishes spoken-language translation, Easy Read and British Sign Language. BSL is a language, not English captions. The BSL Act 2022 concerns England, Scotland and Wales and requires ministerial departments to report on their use of BSL; Northern Ireland has a separate legal context.

For wider inclusive product work, see our assistive-technology and accessibility guide.

Respect Welsh and Other Language Commitments

Evidence of user need should drive language coverage, but statutory commitments may set a floor. GDS translation planning guidance tells government publishers to follow organisational policies, look for user evidence and publish translations as separate accessible pages rather than mixing languages indiscriminately.

In Wales, the Welsh Language (Wales) Measure 2011 enables standards to be imposed through compliance notices. Digital Public Services Wales' 2026 code-of-practice summary recommends bilingual-by-design services for public bodies. Check the organisation's actual notice and contract: a machine-generated Welsh page is not evidence of equal quality, simultaneous availability or compliant service.

Maintain parity logs for changed content. A source release should not quietly make the target version obsolete. Define which changes block publication, which may use a temporary reviewed notice and how users report a language defect.

Protect Content and Personal Information

Before sending text to a vendor, identify controllers, processors, subprocessors, hosting locations, retention, model-training terms and deletion evidence. Minimise inputs; use synthetic test data and redact identifiers where the task does not require them.

If personal information is made available to a recipient outside the UK, assess whether this is a restricted transfer. The ICO's updated international-transfer guide sets out the three-step assessment and relevant safeguards. EU GDPR or other destination rules may apply separately.

Treat uploaded documents and translation memories as untrusted input. They can contain hidden instructions, malicious links or sensitive data. The NCSC's secure AI development guidance covers secure design, development, deployment and operation. Use least privilege, separate customer corpora, restrict tool access, scan files, encrypt data, log administrative actions and test deletion. A translation engine should not browse, send messages or modify production content merely because a source document tells it to.

Our UK AI privacy guide provides the full DPIA, lawful-basis and supplier-review workflow.

Keep Consumer Claims Consistent

Translate the final price, mandatory fees, availability, guarantee, cancellation rights and material limitations—not a shortened marketing approximation. Review generated product descriptions against the approved source and local evidence.

The CMA's unfair commercial practices guidance covers misleading actions, omissions and other prohibited business-to-consumer practices under the Digital Markets, Competition and Consumers Act 2024. The business remains responsible when an AI or supplier produces the words. Keep every locale tied to the same claim register and evidence owner.

For multilingual support journeys, our customer-service AI guide covers escalation, recordkeeping and truthful automation boundaries.

Run a Measurable 90-Day Pilot

Day-90 expansion requires:

Pause a locale when the source version is unknown, a mandatory reviewer is unavailable, critical error recurs, a vendor changes retention or model behaviour without approval, or users cannot reach the human-language alternative.

Good localisation makes the product feel intentionally built for its users. AI can accelerate the draft and the checks, but accountable people still decide what the organisation means—and verify that every user receives it.