AI Translation and Localization

Translation is solved, localization is not

In practice the two words get used interchangeably, but their scope differs. Translation carries a sentence from one language into another. Localization adapts a product, document or campaign to the linguistic, cultural and legal expectations of a given market — the standard definition covers number, date and time formats, currency, sorting order, symbols and colours, text inside images, and differing legal requirements as well.

That difference decides what a machine can be handed. Carrying the sentence across is reliable today. Deciding whether the German page should quote euros, whether to address the reader formally, and whether the date of your holiday campaign even means anything in the target market — those are not translation questions, and the machine does not know it ought to ask them.

Where machine translation stands

At the field's annual shared task, where dozens of systems translate the same texts, in 2025 every external entry but one used a language model rather than a classic translation engine. The best system landed in the winning cluster for 14 language pairs and finished above the human translation in 10 of them. The reverse held too: the human reference translation made the top cluster in only six of the 15 pairs where it was available.

That, however, is true for the large languages. In the same evaluation, the English-to-Maasai direction saw the best open system score 4.8 points out of 100 and the winning closed model 9.8 — the same model that beat human level elsewhere. The gap is not a gradient but a cliff: below a certain volume of training data, quality collapses.

Most languages sit between those two extremes, closer to the thin end than English speakers assume. In the most recent large web corpus, English accounted for 40.45 percent of the text, Spanish for 4.63 and Hungarian for 0.53 — an order of magnitude between two European languages that feel comparably ordinary. This shows up in more than translation quality: text in these languages costs more in tokens, and on the same task the ranking of models differs from the English ranking. Which models actually hold up outside English is its own post: handling smaller languages covers it, and the answer is not the usual order.

What translation carries and localization does not

There are six areas where raw translation reliably produces a bad result, and in all six the error comes from the model not knowing who the text is for.

• Units and numbers. A mile does not turn into a kilometre by itself, and swapping the decimal comma for a decimal point corrupts the figure silently. This is the one error type that costs money immediately.
• Date, time, currency. 03/04 can mean a different day in the target language. With currency the question is not the conversion rate but which currency you should be quoting at all.
• Forms of address. Formal or informal — the model decides sentence by sentence and is not consistent unless you tell it.
• Legal and official register. Here mistakes are most expensive, because the error does not sound like one. An English trial is a trial period in a contract and a court proceeding in a filing — the surrounding text does not help the model choose.
• Idiom, humour, wordplay. Whatever can be taken literally will be taken literally. A translated idiom usually still parses, it is simply not funny, and occasionally it is unsettling.
• Text inside images. The most commonly forgotten element: text baked into banners, diagrams and screenshots is not part of the translatable material, so it stays in the source language, and nobody notices until launch.

The form of address that breaks most often

In most European languages other than English, formal versus informal address is not a style question but the text's very first decision, and it has to hold all the way through. Machine translation makes that decision sentence by sentence, so both show up in the same letter. When the source is English the model has no signal at all to work from: English you means both.

It is worth knowing that this is not only a weakness of the models. One of the market's main translation engines has a dedicated formality switch, but it supports a specific list of target languages — Spanish, German, French, Polish, Japanese and a handful of others — and everything outside that list is left to you, whether by prompt or by a pass afterwards. The field has moved on this: at the 2026 shared task, ignoring an instruction about the form of address counts as a translation error.

The workflow that holds

Five steps, and the order matters more than the choice of tool.

1. Write the localization sheet. Form of address, currency, date format, units, banned words, and the list of brand names that are not translated. It is written once and handed over with every job.
2. Translate by section, not by sentence. Sentence-by-sentence translation loses the connections in the text; the whole document at once degrades with length. The paragraph or the section is the right unit.
3. Supply terminology separately. Give a word list of the technical terms with the translation you want. This is the single intervention that measurably improves output on every model.
4. Post-edit, do not proofread. Raw machine output is not a deliverable — the professional standard says so explicitly. In practice the pass runs from the word to the paragraph, not a tick at the end.
5. Have a native speaker read the target. At minimum the opening and closing sections, the headings and every number. This is the twenty percent that catches eighty percent of the errors.

How to measure quality

Treat automatic scores carefully. In the field's 2025 evaluation, scoring done by a large language model correlated well with human judgement at system level, while at sentence level the older, simpler metrics — including the character-based chrF — beat the modern learned scorers. Ranking systems and judging a single sentence are not the same tool.

The most instructive case comes from the same evaluation: one system won nearly every automatic ranking, then landed well below the leaders once humans assessed it. If you had gone by the score, that is precisely the system you would have picked.

One thing follows for practice: twenty sentences from your own material, translated and checked by a human, are worth more than any public score. The same twenty sentences across every model, and the decision is yours.

What this means for prices

Industry survey data puts the share of independent translators' work touched by some form of automated translation at 63 percent, the discount applied to post-editing at around 20 percent on average, and the share of classic human translation in agency revenue down from 37 to 29 percent in a single year. Over the same period, adoption of generative AI in corporate language departments rose from 30 to 64 percent.

What that does not mean is that human work has left the process. Market analysis describes the dominant model as expert-in-the-loop, with human oversight as "a non-negotiable safeguard". Clients expect cost reductions of up to 75 percent; the industry can deliver a fraction of that while holding quality. If the decision is yours, the real question is not machine versus human — it is where your text sits on the scale between an error being awkward and an error being expensive.

Typical failures worth knowing in advance

These are error types that have actually happened, not theoretical risks.

• The literal technical term. An English term turns into a similar-sounding, nonsensical word in the target language. Only a native speaker catches it, because it is grammatically flawless.
• The look-alike word. Pairs like week and milk cause a silent swap: the sentence still makes sense, it just means something else.
• The legal ambiguity. A trial period becomes a criminal proceeding, a notice period becomes a dismissal. If your text is a contract or a set of terms, do not leave that section to a machine.
• The loss of consistency. The same product or feature name appears three different ways in one document. Without a glossary this is close to guaranteed.

What to take away

Carrying the sentence across is solved; deciding what the text should be in the target market is not. Write a localization sheet with the form of address, currency, formats and untranslated names, supply a terminology list, and have a native speaker read the target. Outside the handful of languages the tools support directly, fixing the form of address is a separate job. And measure quality on twenty sentences of your own, not on a public score.