content-locale-humanize-ko · diff
git:20260918.40e721d to git:20261013.a69e0bb
18 added, 9 removed. Audit A to A.
---
name: content-locale-humanize-ko
title: "Korean Native-Fluency Calibration"
description: "Per-language calibration anchors for detecting AI-slop and translationese in Korean (ko) target text; auto-loads alongside content-locale-humanize when the target language is Korean."
license: Apache-2.0
- compatibility: "Extends content-locale-humanize's AI-Slop & Translationese dimension. Load both together — this file has no rubric of its own."
+ compatibility: "Calibration data only; no tools required. Network access helps for usage checks."
domains: content
rules:
- content(korean)
- match(\bko-KR\b)
- match(\bKorean\b)
- match(한국어)
- semantic(check if this Korean translation sounds native)
- semantic(eliminate translationese from this Korean text)
---
## Overview
- Sourced calibration anchors for Korean, feeding `content-locale-humanize`'s AI-Slop & Translationese dimension. This is a calibration aid, not the checklist — reason natively beyond it (see the core skill's "why structure, not word lists" section).
+ Calibration anchors and current norms for Korean (ko-KR), feeding content-locale-humanize's dimensions. A calibration aid, not the checklist; reason natively beyond it.
- ## Instructions
+ ## Rules
- Sourcing confidence: moderate, but with one structurally important, well-reasoned finding below.
+ ### AI tells
- Korean makes up roughly 1% of typical LLM training corpora versus roughly 60% English — meaning Korean output starts from a shakier native-fluency baseline than most other languages checked here. Weight interference and awkward-register findings a little more heavily for Korean; the model has comparatively less native Korean signal to draw on, so translationese bleeds through more easily.
+ Sourcing confidence: moderate.
- Connector pileup: 그리고 (and), 하지만 (but), 그래서 (so) — mechanically chaining sentences instead of letting the logic connect on its own, a substitute for genuine narrative flow.
+ - Korean is a small share of typical LLM training data compared with English, so interference bleeds through more easily; weight interference and register findings slightly higher.
+ - Connector pileup: 그리고, 하지만, 그래서 chaining sentences mechanically instead of letting the logic connect.
+ - Adverb-modified generic verbs (-게, -이, -히) where a concrete verb would carry the meaning.
+ - 중요한 / 중요하다 measured as disproportionately overused against native baselines.
- Overused descriptive-adverb endings: -게, -이, -히 — these tell rather than show; native Korean writing favors more concrete verbs over adverb-modified generic ones.
+ ### Translationese (번역투)
- Frequency-outlier word: 중요한 (important) and its English cognate "significant" — both measured as disproportionately overused versus native baselines.
+ - "~에 의해" passives, "~을/를 가지다" for "have", "~에 있어서", "~하는 것이 필요하다", and "~를 통해" everywhere.
+ - Pronouns 그/그녀 and plural -들 where Korean leaves them implicit; stacked -적 adjectives.
+ ### Current norms
+
+ - Speech level: 합니다체 for formal documents and notices, 해요체 for most consumer apps and UI; pick one and hold it.
+ - Word spacing (띄어쓰기) and loanword spelling (외래어 표기법: 콘텐츠, 메시지) follow the National Institute of Korean Language; spacing errors read as machine output immediately.
+
## References
- rebrandb.com, zdnet.co.kr, curious-500.com.
+ Reviewed 2026-10-12. AI tells: rebrandb.com, zdnet.co.kr, curious-500.com. Norms: [국립국어원](https://www.korean.go.kr/).