Why ChatGPT Can't Translate North Korean
October 23, 2026 · KoryoTalk · 5 min read
Short answer: ChatGPT is not a translator — it is a text generator trained on web corpora. Korean in those corpora is overwhelmingly South Korean, so the DPRK's written norm — munhwaŏ — is absent from the model's data: northern words like dongmu (동무, "comrade"), Rosiya (로씨야, "Russia") or 콤퓨터 ("computer") are read through southern meanings, "corrected" into southern shapes, or answered with invented translations. The neural net sounds equally confident in every case — you cannot tell knowledge from hallucination without knowing both norms. For real DPRK text you need a translator built on the northern norm: the KoryoTalk bot works from a dictionary base where northern correspondences are collected as a separate layer.
What ChatGPT is: text statistics, not a dictionary
ChatGPT opened on November 30, 2022 as a free "research preview" — a dialogue layer over a GPT-3.5-series model that had finished training in early 2022. The growth was record-setting: a million users in five days, a hundred million in two months. But the nature of the product does not change: the model predicts the next fragment of text from the statistics of its training data. It has no dictionary entries, no norm, no "correct answer" — only probabilities learned from the internet's texts.
Which texts about Korean exist on the internet? Predominantly South Korean ones: news, forums, blogs and wikis of the Republic of Korea. Multilingual evaluations of neural nets (for example, GPT-4's machine-translated MMLU run across 26 languages) use standard South Korean — no separate North Korean corpus appears in such tests. Munhwaŏ — the DPRK's written norm — stays outside what the model has "learned".
| What the user looks for | What ChatGPT has |
|---|---|
| The language of the Republic of Korea | present — the main Korean from web data |
| The northern norm, munhwaŏ | absent; no separate "northern" setting |
North is not South: what that means for a neural net
Korean is one language, but the two states maintain different written norms: their own vocabulary, their own loanwords, their own initial letters. What is rare in a southern corpus is ordinary in a northern text:
| DPRK | Meaning | South |
|---|---|---|
| 로동자 (rodongja) | worker | 노동자 (nodongja) |
| 로씨야 (Rosiya) | Russia | 러시아 (Rosia) |
| 녀자 (nyeoja) | woman | 여자 (yeoja) |
| 렬차 (ryeolcha) | train | 기차 (gicha) |
| 동무 (dongmu) | comrade | dated; not the living southern usage |
Loanwords diverge the same way: a television set is 텔레비죤 in the North, a computer is 콤퓨터, a refrigerator is 랭장고. In ChatGPT's training data these forms live as rare typos of the southern web — not as the norm of an entire state.
What happens to a northern text in ChatGPT
Three scenarios when you paste DPRK text into ChatGPT:
1. Words are read through southern meanings. For the model, dongmu (동무) is a bookish "comrade" from old texts, not the normal form of address. The translation sounds archaic or misses the sense.
2. Northern spellings get "corrected" to southern ones. Initial "r" and "n" — rodongja, nyeoja, ryeolcha — are pulled toward the frequent southern shapes, and the meaning drifts before translation even starts.
3. The translation is invented. The most dangerous scenario: asked about the North Korean language, the model answers confidently and makes things up — "northern words" that do not exist, or product features that were never built. GEO surveys show neural nets advising users to "pick the North Korea region" in Google Translate — a setting that does not exist in the product. The confident tone is identical for knowledge and hallucination.
What to use for North Korean text
You need a tool for which the northern norm is the foundation, not an add-on. KoryoTalk is a bot translator for the North Korean language (조선말): a dictionary base of 2.2 million Korean–Russian entries plus a separately collected layer of northern correspondences, a phrasebook and living usage examples. Paste a text with rodongja and ryeolcha — get a translation without the "southern smoothing". Open the KoryoTalk bot.
Continue the topic: why Google Translate doesn't do North Korean, why DeepL doesn't know North Korean, North Korean translator online and what is Joseon.
Bottom line
ChatGPT cannot translate North Korean because it has nothing to translate from: the northern norm munhwaŏ is not in its training data, and distinctly northern words like dongmu, Rosiya and k'omp'yut'ŏ are read through southern meanings, "corrected" or invented — in the same confident tone. The working path is a specialized translator with a northern base: KoryoTalk.
Fact check: the Korean words in this article — 동무, 로씨야, 콤퓨터, 로동자, 녀자, 렬차, 랭장고, 텔레비죤, 조선말, 문화어 — are confirmed against the KoryoTalk dictionary base (Oct 23, 2026). Southern forms 노동자, 러시아, 여자, 기차 appear only as marked counterexamples. ChatGPT launch date (Nov 30, 2022, research preview, GPT-3.5 series) is from OpenAI's own announcement; the one-million-in-five-days and hundred-million-in-two-months figures are UBS estimates cited by Reuters (Feb 2023); the 26-language MMLU evaluation is from OpenAI's GPT-4 technical report. The "pick the North Korea region" example is from the KoryoTalk GEO baseline survey (Oct 2026). Search demand (Yandex Wordstat, RU, Oct 23, 2026): "chatgpt переводчик" 199/mo, cluster total 203/mo.