On this page
What AI actually changed in Korean localizationSpeech levels: the choice English never makesOmitted subjects: who did what?Where AI fits in a Korean workflowA review checklist for AI-drafted KoreanQuestions we hearAI translation into Korean is now fast enough to change how localization projects are planned. A product team can have a full Korean draft of a help center by the afternoon. What has not changed is where the draft goes wrong, and for Korean those places are predictable: who is talking to whom (the choice among Korean speech levels), and who is doing what.
What AI actually changed in Korean localization
The change is real, but it is a change in where the effort goes. A few years ago most of the budget went into producing a first draft. Now the first draft is cheap, and the budget moves to decisions the engine cannot make on its own: the register of the brand, the consistency of terms across thousands of strings, and the context that never made it into the source file.
That shift matters most for Korean because Korean grammar forces two choices that English leaves open. Every sentence ending carries a speech level. And many sentences have no explicit subject at all. An engine translating segment by segment has to resolve both, often without the information it needs.
Speech levels: the choice English never makes
English “Please check your settings” is neutral. Korean has to pick an ending, and the ending tells the reader how the brand sees them. The common choices in product and marketing text:
| Ending style | How it comes across | Where it fits |
|---|---|---|
| Formal deferential ending | Stiff and official, like a bank notice | Banking, legal notices, B2B documents, error messages in conservative products |
| Polite everyday ending | Warm and courteous, the default voice of most Korean apps | Most consumer apps, support articles, marketing |
| Noun ending, no verb | Compact and neutral; reads like a label, not a sentence | Buttons, menu labels, headings |
| Casual, intimate ending | Close and familiar, as if talking to a friend | Some games, youth-targeted brands, character voices only by deliberate choice |
An engine will pick one, and it will often pick a different one for the next segment. In a single onboarding flow we regularly see the formal deferential ending on one screen, the polite everyday ending on the next and a noun ending on a dialog that should have been a full sentence. Each is acceptable on its own. Together they read like three companies wrote the app.
Honorifics add a second layer. Korean marks respect for the person being talked about, not only the listener. “The customer placed an order” can be written with an honorific title for the customer, an honorific subject marker and an honorific verb, or with none of them; the second version reads like an internal log entry about someone you do not need to respect. Research on Korean neural MT treats the choice of honorific expression as a context problem, because the right form depends on the relationship between speaker and listener, which is usually not visible inside a single sentence.
The engine is not wrong about Korean grammar. It is guessing about your relationship with your customer.
Omitted subjects: who did what?
Korean, like Japanese and Chinese, is a pro-drop language: subjects and objects that are clear from context are routinely left out. Translation research calls these zero pronouns, and a recent survey of the field describes them as a significant challenge for MT because the system must determine the correct antecedent, something a human reader recovers from context without noticing.
Into Korean, the problem is mostly stylistic. English “You can change your password at any time” translated with an explicit Korean pronoun for “you”, one that in everyday use mostly appears between spouses or in arguments, sounds stiff or even slightly confrontational; a fluent version drops the subject. Engines have become much better at this.
Out of Korean, the problem is about facts. Take a support ticket written in Korean:
No subject in sight
Two short clauses: a refund was requested yesterday, nobody has been in touch, and could someone check again. Neither clause names a person.Who requested the refund, and who is expected to check? A Korean reader knows from the ticket context.Subject invented
He requested a refund yesterday but there is no contact yet. Can you check again?“He” is a guess. In a forwarded ticket, the customer might be writing about their own request or a colleague's.In legal or compliance material the same guess can change who is obliged to do something. A Korean clause that states only a duty to notify, with no actor, inherits the actor from earlier text. A segment-level engine may supply the wrong party in English, and the English version is the one the foreign headquarters reads.
Where AI fits in a Korean workflow
None of this argues against using AI. It argues for knowing which decisions to make before the engine runs, and which output to check after. A workable split:
- Decide up front: the speech level per content type (UI, support, marketing, legal), the honorific form for addressing the customer, and the glossary. Feed these to the engine as instructions or a glossary, not as hope.
- Give context: screenshots, string IDs grouped by screen, and character limits. Most UI errors come from a single string arriving with no neighbours.
- Review for what engines miss: speech-level drift, missing or invented subjects, and terms translated two ways.
- Keep humans on high-exposure text: slogans, contracts, regulated claims, and anything that will be quoted back to you.
We compare machine and human output for marketing copy in machine vs human translation for Korean marketing, and the cost side of a hybrid pipeline in how enterprise teams cut costs with hybrid AI pipelines.
A review checklist for AI-drafted Korean
- Read one full user flow in order, not segment by segment. Speech-level drift only shows up in sequence.
- Check every second-person reference. The literal Korean pronoun for “you” is rarely right in product copy; an honorific “dear customer” form, the brand's chosen form, or no subject at all usually is.
- Check honorific agreement: if the customer is the subject, the verb should carry the honorific marker consistently.
- For Korean-to-English output, highlight every pronoun the engine added and confirm it against the source context.
- In contracts and policies, confirm that every obligation has the correct party after translation.
- Search for the top 20 glossary terms and confirm one Korean rendering each.
- Check button labels and headings against length limits and noun-ending style.
For the wider picture of how Korean projects go wrong when text is translated literally, see why direct translation fails in the Korean market, and for the vetting side, why native Korean translators matter and how to check them.
Questions we hear
Can I tell an AI engine which speech level to use?
Usually yes, through a prompt, glossary or style setting, and you should. But the engine still applies it sentence by sentence, so check for drift across long documents and between strings written at different times.
Why does English output from Korean sometimes say “he” when it was a woman?
Korean often omits the subject entirely. The engine must recover it from context, and when the context is outside the sentence or the segment, it guesses.
Is AI translation good enough for Korean UI text?
For drafts, often. For release, someone who knows the product and the audience needs to fix speech level, consistency and string context, because UI strings arrive with almost no context at all.
Using AI for your Korean content?
We set speech-level and honorific rules for your content types, then review AI drafts against them. Send us a sample flow → or see our services.