imnotai
Stripping the AI tells out of Korean prose without touching a single fact — 70 catalogued patterns, a change-rate hard stop, and an honest story about its own false positives.
Korean text written by ChatGPT, Claude, or Gemini reads off in a way native speakers clock instantly and struggle to name. imnotai names it: 10 categories and 70 sub-patterns, catalogued, severity-ranked, and rewritten span by span — while the facts, numbers, and quotes stay byte-identical.
It comes in two forms. imnotai.kr runs in a browser with nothing to install and nothing to sign up for, for people who can’t put a CLI on the machine they’re writing on. im-not-ai (“Humanize KR”) is the same taxonomy as a CLI skill, MIT-licensed, and the version you want if you’re already living in a terminal.
Why a Korean-specific tool
QuillBot, Hix, and the other English humanizers are weak here, because Korean AI tells mostly aren’t stylistic — they’re English translationese leaking through:
| The tell | What it should have been |
|---|---|
| AI 기술을 통해 효율을 높일 수 있다 | AI로 효율을 높인다 |
| 이에 있어서 중요한 점은 | 여기서 중요한 건 |
| ~에 의해 생성된 | ~가 만든 |
| 결론적으로, 이는 시사하는 바가 크다 | (delete it) |
The full catalogue runs from translationese (A) through mechanical parallelism — 첫째/둘째/셋째 (C), AI stock phrases (D), uniform sentence rhythm (E), hedging pile-ups (G), sentence-initial connective spam (H), to bold-and-em-dash decoration (J). Each pattern carries a severity: S1 means one occurrence is already a giveaway, S2 tolerates one or two, S3 only matters when it stacks with something else.
The four rules it won’t break
- Meaning is invariant — facts, claims, figures, proper nouns, and direct quotes are preserved 100%.
- Evidence-based edits — only detected spans get touched; undetected text is left alone.
- Genre holds — a column doesn’t come back as literature, a report doesn’t come back as an essay.
- No over-editing — past a 30% change rate it warns, past 50% it hard-stops.
That last one is enforced by a script with an exit code (verify_change_rate.py), not by
asking a model to be careful, which is the right place to put a guardrail. There’s also
an explicit do-not-touch list: numbers, units, dates, proper nouns, anything inside
quotation marks, statutory text, and unavoidable academic terminology.
Install the CLI version
/plugin marketplace add epoko77-ai/im-not-ai
/plugin install humanize-korean@im-not-ai
Then /humanize-korean in a new session, or just ask in plain Korean — “이 글 AI 티
없애줘”, “번역투 제거” and similar phrasings all trigger it. GitHub Copilot CLI has the
same marketplace path (copilot plugin install humanize-korean@im-not-ai), and there’s a
./install.sh for a cloned checkout.
One asymmetry worth knowing before you pick a host: Copilot CLI, Codex CLI, and Gemini CLI get the single-call path only. The multi-call route with a separate diagnostic pass and a finalizer is Claude Code exclusive.
Cost comes from call count, not model choice
The interesting engineering here is the router. A scoring script grades the input first and deterministically picks one of three paths:
| Path | LLM calls | When |
|---|---|---|
| light | 1 | Already well-written — may exit early with “this is fine” |
| standard | 2 | An ordinary AI draft: diagnose once, rewrite once |
| heavy | 3+ | Dense slop, over 15,000 characters, or you asked for an audit trail |
Savings come from making fewer calls, not from downgrading the model — which stays your choice. The measurement that drove the design: a 10,000-character piece run as 7 chunked calls burned 610K tokens; the same piece as a single call took 134K at equal quality. Chunking reloads the rulebook and the diagnosis for every chunk, and that overhead eats the entire benefit. Chunking is now heavy-path only.
The part that earned my trust
In August 2026 the data-quality firm Pebblous took the public code apart. They confirmed the citation-tampering and over-editing guards worked — and then handed over a reproducible counterexample: a human essay from 2020, written before ChatGPT existed, scored 6/6 maximum AI risk. The cause was one author’s personal comma habit (83.3% of connective endings followed by a comma) sending a single pattern’s z-score through the roof.
Re-testing against 532 clean human documents turned a one-off into a system defect:
| Detector version | Human text misjudged “high” | medium | low |
|---|---|---|---|
| Before | 285 (53.6%) | 171 | 76 |
| After the family cap + two-independent-signal rule | 0 | 4 | 528 |
| After genre-aware cells as well | 0 | 0 | 532 (100%) |
More than half of human writing was being flagged, and the project published the number rather than quietly patching it. That’s the response you want from a tool that renders judgements about whether a person wrote something.
The lesson generalises past this repo: no single stylistic metric is evidence. Anyone using an AI detector as a verdict on a student, a colleague, or a job applicant should read that teardown first.
What it isn’t
The author is direct about this, and it’s worth repeating: this is not an “undetectable AI” bypass. The goal is Korean writing quality — the tool improves prose that happens to have been drafted by a model, and it is explicitly not a guarantee for academic submission or journalistic integrity. If your institution requires disclosure, run this and disclose anyway.
Pairs naturally with Markdown for drafting and Obsidian for keeping what you wrote.
Next
When the document to produce is a whole book → Book Writer
Last updated 10 Sep 2026, 00:00 UTC.