AI & Development
AI has transformed app localization - but using it well requires understanding its strengths, its failure modes, and how to build a quality review process
App localization has traditionally been expensive and slow: managing translation vendors, coordinating review cycles, handling technical constraints like string length limits, and maintaining consistency across app versions. AI-powered translation and localization has dramatically accelerated this process. Modern LLMs produce translations that are significantly better than older machine translation systems, understand context across strings, and can be guided by style guides and terminology glossaries. But using AI for localization well requires understanding where it excels and where human expertise remains essential.
General UI text - buttons, labels, menu items, error messages - is where AI localization performs best. These strings are short, context-independent enough to translate reliably in isolation, and forgiving of slight imperfections. An AI-translated "Save" button is indistinguishable from a human-translated one in almost all languages. An AI-translated error message like "Please enter a valid email address" is accurate and natural across dozens of languages without requiring human review in most cases.
AI also handles structural and formatting constraints well when told about them. A prompt that includes "This string will appear in a mobile button with a maximum width of 12 characters in English. Translate to French, but keep the translation short enough to fit in a similar space" produces usable translations more reliably than raw machine translation that ignores UI constraints.
Idiomatic language is the primary weakness. Expressions, metaphors, humor, and culturally-specific references rarely translate well with AI. An AI that translates "you're in the driver's seat" literally into a language where that phrase has no meaning produces a confusing localization. App descriptions, marketing copy, and any text that relies on language-specific rhetoric requires human review by a native speaker who understands the cultural context.
Technical terminology in specialized domains is another weakness. Medical, legal, financial, and scientific terminology has precise meanings in every language, and correct translation requires domain expertise that general-purpose LLMs do not reliably have. Providing a terminology glossary in the prompt - "translate using these approved terms: [glossary]" - significantly improves technical translation accuracy, but does not eliminate the need for expert review in high-stakes domains.
Translating strings in isolation is the most common failure mode in AI localization. A string like "tap to continue" might be clear as an isolated string, but its translation depends on what the user is continuing from, who the audience is, and what tone the app uses throughout. Providing surrounding context in the translation prompt - a description of where the string appears, what precedes it in the flow, and the app's overall tone guidelines - produces substantially better translations.
Context batching is a practical technique: rather than translating strings one at a time, group related strings (all strings from the onboarding flow, all strings from the settings screen) and translate them together, providing the screen description as shared context. The model can then use each string as context for others in the same group, producing more consistent and contextually appropriate translations.
A practical AI localization workflow: AI translation first, followed by a native-speaker review of a sampled subset (typically 10-20% of strings), with full review for marketing copy and any user-facing content that reflects the brand voice. Track which types of strings the AI handles reliably (UI labels, short messages) and which require review (long descriptions, idiomatic expressions). Over time, you accumulate data that lets you intelligently route strings to AI-only or AI-plus-review workflows based on predicted error rate.
Maintaining a translation memory - a database of previously approved human translations - allows you to enforce consistency across versions. When a string has been previously translated and reviewed, use the approved translation rather than re-translating with AI. Translation memory reduces both cost and inconsistency.
The standard automated metric for translation quality is BLEU score, which measures n-gram overlap between the AI translation and a reference translation. BLEU is useful for trending quality over time and catching significant regressions, but it correlates imperfectly with human judgment on what "sounds natural" in a given language. Human evaluation remains necessary for localizations where naturalness matters to the user experience, and a small panel of native speakers evaluating a sample of translations is more informative than BLEU scores alone.