How AI translation fails in localization files — a mechanical look

AI translation systems are trained to produce fluent, natural-sounding sentences in the target language. That is exactly what makes their mistakes dangerous in a localization file: a broken translation usually still reads as a well-formed sentence, so nothing about the output looks wrong on its own. You have to know the specific ways it can fail and check for them deliberately.

None of what follows is about any particular product — these are structural failure modes that follow from what these systems are optimizing for (fluency and plausibility of the target sentence) versus what a localization file actually needs (exact preservation of everything that is not prose).

Placeholders and variables

Localization strings carry placeholders like {playerName}, %d, or {0} that get substituted at runtime. A model translating the surrounding sentence has no obligation to treat the placeholder as an opaque token — it can reorder it into a grammatically more natural position for the target language, alter its spelling or casing, or in rarer cases drop it entirely if it seems redundant to the sentence's meaning.

source: "You found {itemName} x{count}!"
risk:   "{itemName}を{count}個見つけた!" -> placeholder order flipped, or
        one placeholder silently dropped in a longer sentence

Markup and tags

Rich text tags (bold, color, line-break tags) embedded in a string are not part of the sentence's meaning, but they sit inside the text the model is rewriting. A tag can be moved to a position that no longer wraps the intended word, have its opening or closing half dropped, or have its attribute value altered — all of which are invisible if you only read the rendered translation as prose and don't check the tag structure.

Formatting and whitespace

Leading or trailing whitespace used for UI padding, intentional line breaks inside multi-line dialogue, non-breaking spaces, and consistent punctuation width all fall outside 'the meaning of the sentence' and are exactly the kind of detail a model optimizing for natural output is not obligated to preserve.

Terminology drift across a long file

A single sentence translated in isolation has no memory of how the same term was translated three thousand lines earlier. Over a long file, a term that should stay fixed — an item name, a stat name, a recurring piece of UI vocabulary — can legitimately shift to a different but equally valid translation partway through, simply because each segment was generated independently. The result reads fine sentence by sentence and is inconsistent as a whole.

Fluent but context-blind

A model translating a string in isolation does not know if it is a button label, a full sentence of dialogue, or a tooltip — and that context changes the correct grammatical form in many languages. A short imperative UI label can come back as a full, grammatically complete sentence; a line of first-person dialogue can come back in a neutral, narrator-like register. The translation is fluent and wrong for its slot.

Untranslated passthrough and register

Some lines can pass through unchanged — copied instead of translated — and still look plausible if the source text happens to resemble valid target-language text (short strings, numbers-and-symbols strings, or loanwords are the common case). Separately, in languages with grammaticalized politeness levels, such as Japanese, a model can slip between formal and casual register line to line, or give a character a level of politeness inconsistent with the personality established elsewhere in the same file.

What is checkable by machine, and what is not

This is the practical line to draw. Placeholder presence, count, and structural tag balance are mechanically verifiable: you can compare the placeholder set in source and target line by line without understanding either language. Formatting and whitespace differences are similarly comparable as strings.

Terminology consistency, tone and register, and whether a translation fits its UI slot are not mechanically decidable from the string alone — they need either a maintained glossary to check against, or a human reader who understands the target language and the context the line appears in. Automated checks narrow down where to look; they do not replace that read.

• Mechanically detectable: missing/altered placeholders, broken or unbalanced tags, whitespace and line-break differences, exact untranslated duplicates of source
• Needs human or glossary-based review: terminology drift, register/politeness consistency, whether tone fits a UI slot, whether meaning is actually preserved