“Good enough” is the most dangerous phrase in localization, because the answer depends entirely on what you are about to do with the text. A traffic-light test for languages that machines handle unevenly.
THE SHORT VERSION
- Machine translation quality is not equal across languages. Languages with less digital text tend to get weaker, less consistent output.
- The errors are quiet: fluent sentences that say something slightly different from the source.
- Use a traffic-light test: GREEN for gist, AMBER for post-edit, RED for human-first.
- The mistake is not using machines. It is using one standard for all three colors.
“Good enough” is the most dangerous phrase in localization. It sounds like a decision, but it is really an unanswered question: good enough for what?
A machine-translated paragraph that is good enough to help a manager understand an incoming email may be nowhere near good enough to appear on a product page, in a contract or on a medicine box. The words might be the same. The consequences are not.
Not all languages are treated equally
Machine translation has improved enormously for widely used languages, where systems have seen vast amounts of text to learn from. But quality is not equal across languages. Languages with less digital text to learn from, often called low-resource languages, tend to get weaker and less consistent output.
More importantly, the errors are quieter. In a language where the machine is weak, the output can still look polished. It reads like a fluent sentence. It just says something slightly different from the source: a negation dropped, a number moved, a formal tone replaced with a casual one, a technical term swapped for an everyday word. Unless the reader speaks both languages well, nothing signals that anything went wrong.
“The most dangerous machine output is not bad language. It is confident, fluent language that contains a quiet mistake.”
What makes these three languages interesting
Georgian and Armenian each use their own unique alphabet, found almost nowhere else. That affects more than translation: fonts, layouts, text fields and PDF exports can all break if the writing system was an afterthought. On the language side, both have rich word structure, where a small change at the end of a word can shift who did what to whom.
Armenian adds a decision that no machine can make on its own: Eastern Armenian, the standard used in Armenia, or Western Armenian, the standard of much of the diaspora. A system will often pick one without telling you which.
Kazakh raises its own questions. It is built from suffixes that carry a great deal of meaning, it uses Cyrillic with nine letters beyond the Russian alphabet, and the question of whether and when it will move to Latin script has been discussed and revised more than once. Terminology in specialized fields is also still settling, which means a machine may produce a word that is technically understandable but not the one a professional in the field would use.
A traffic-light test we use
The most useful way to think about this is not “machine or human” but “how much risk is attached to this text”. We use a simple traffic-light test.
GREEN: internal gist. You only need to understand what a document says, for example a competitor’s press release or a customer’s long email. Machine translation is fine. Nobody outside your team will read the result, and a human can follow up if something looks important.
AMBER: public-facing but low-risk. A support article, a help-center page or a blog post that nobody will make a decision on. Machine translation plus post-editing by a native linguist who knows the subject. Post-editing is not proofreading: the editor has to read against the source, not just check that the target sounds right.
RED: contracts, medical or safety information, brand and marketing copy, user interface text. Human-first, with machine output used only as a draft if at all. In a RED text, a quiet mistake can cost a lawsuit, a recall or a brand’s reputation.
The mistake is not using machines. The mistake is using one standard for all three colors.
Questions to ask any vendor
- Which of my content types do you treat as GREEN, AMBER or RED, and why?
- Who post-edits for my language, and are they native speakers with subject knowledge?
- How do you decide between Eastern and Western Armenian, or between scripts, before you start?
- Where does my content go when machine translation is used, and who can access it?
- What is checked by a person before delivery, and by whom?
Where would you draw the line?
These are general observations about how the technology tends to behave, and they will keep changing as systems improve. That is exactly why the line should be a decision your team makes deliberately, rather than a default your vendor makes for you.
So we would like to hear where you would draw it for these languages. And translators: what is the strangest machine error you have had to fix?
Tell us in the comments. We reply to everyone.
Language Shift by Transinity: AI, localization, careers and culture from inside the language industry. Follow Transinity on LinkedIn and join the conversation.