On this page
How machine output failsContent type by content typeWhat large institutions actually doChecking machine output efficientlyPractical rules for a Korean rolloutA team machine-translates its whole help center into Korean over a weekend. Most articles are fine. One, about account deletion, tells Korean users that data "may not be recovered" in a way that reads as "might not be", rather than "cannot be". Nobody notices until a user asks support why their deleted data is gone.
The useful question is not whether machine translation is good. It is which content it is good at. The research community's own title for its 2024 evaluation campaign sums up the state of the art: "The LLM Era Is Here but MT Is Not Solved Yet." Systems are fluent. They are not reliable in every sentence. Whether that matters depends entirely on what the sentence is for.
How machine output fails
Average quality is high, so the failures that matter are the rare, confident ones. A study of neural machine translation by Guerreiro, Voita and Martins annotated more than 3,400 sentences for critical errors and hallucinations, output that is fluent but detached from the source. These errors are hard to spot precisely because the sentence reads well. A reviewer skimming Korean for fluency will pass it.
Korean adds its own pressure points. Korean routinely drops subjects that English states, so the machine must decide who did what; it marks politeness on every sentence, so tone can drift between paragraphs; and many short English UI words are ambiguous without context.
"May not"
EN: Deleted data may not be recovered.Risk: the engine picks the "possibility" reading, so the Korean says recovery might failIntended: recovery is impossible, a "cannot" statementIllustrative. English "may not" allows both readings; Korean forces a choice, and only the writer knows which was meant.No context
EN: PostRisk: Korean has separate words for postal mail, a published post and the act of publishingIllustrative. A button, a noun in a menu and a postal address field all need different Korean, and the string alone does not say which.The missing clause
EN: Refunds are available within 14 days, except for digital items that have been downloaded.Risky output drops the "except" clause and reads fluently.Illustrative. Omissions are the error a fluency check misses.Content type by content type
| Content | Where MT holds up | Where it breaks | Sensible default |
|---|---|---|---|
| Inbound reading (emails, competitor pages) | Getting the gist quickly | When the gist is quoted back as fact | Raw MT, never republished |
| Support replies | Routine answers, known topics | Refunds, account security, angry customers | MT with agent review; templates translated by humans |
| Internal documents | Most of it | HR policies, anything legal | Post-edited MT |
| Product UI | Long descriptive strings | Short strings without context, variables, speech level | Human translation with screenshots |
| Manuals and technical docs | Repetitive procedural text | Terminology consistency, units, safety warnings | MT plus full post-editing against a termbase; warnings by humans |
| Marketing | Rough drafts for ideas | Tone, wordplay, cultural fit | Transcreation from a brief |
| Contracts, labels, medical, patents | Orientation for reviewers | One error can change rights or safety | Human translation with revision and expert review |
The split between "post-edited" and "human" is described in more detail in our machine translation post-editing guidelines, and the levels behind this table in fit-for-purpose quality levels.
What large institutions actually do
The European Commission runs one of the largest translation operations in the world and its own machine translation service, eTranslation, which it says is used by EU staff and other eligible users to produce raw machine translations. Its translators can use machine output as an optional resource, its in-house staff check contractors' translations, and it describes translation memories as crucial for consistent terminology in legislation. In other words: heavy use of MT, and humans accountable for what is published.
Some regulators write this into rules. Under the US Section 1557 regulation for covered health programs, if machine translation is used where the text is critical to the rights, benefits or meaningful access of a person with limited English proficiency, where accuracy is essential, or where the source contains complex, non-literal or technical language, the translation must be reviewed by a qualified human translator. It is a US health rule, not a Korean one, but it is a clear statement of where the line sits.
Checking machine output efficiently
When post-editing is the right level, the review should target the errors machines actually make, not re-read for style. In Korean, that means a short list of checks that catch most of the damage:
- Negations and modals. Search the source for not, never, except, unless, may, must and shall, and confirm each one survived with the right force.
- Numbers, dates and units. Compare digit by digit. Watch conversions into the Korean units of ten thousand and one hundred million, which machines and humans both get wrong.
- Named entities. Product names, company names and people's names should match the termbase, not whatever the engine chose this time.
- Sentence count. If the source has twelve sentences and the target reads as ten, something was merged or dropped.
- Subject. Where Korean omits the subject, make sure the reader can still tell whether "you", "we" or "the provider" is acting.
These checks take minutes per page and catch the errors that a fluent read misses. They do not make raw output safe for regulated content; they make post-editing honest for the content where it belongs.
Practical rules for a Korean rollout
- Classify content before translating it, not after something goes wrong.
- Never publish raw MT under your brand in a customer-facing channel.
- Give post-editors the source, the termbase and screenshots; check omissions against the source, not fluency alone.
- Keep a human translator on anything with legal, financial, safety or patent consequences.
- Track errors found after release by content type, and move content up a level when errors recur.
For the patent-specific version of this argument, see why AI and machine translation are risky for patent documents; for regulated sectors, machine translation in healthcare and legal work; and for the operating model that combines both, combining AI and human translation.
Deciding what can go through MT?
We help teams classify Korean content by risk and then handle the parts that need a human. Tell us what you translate →
Further reading: Global Communication: The Power of Translation Devices in 2025