Korean AI detector
What the Korean AI detector gives you back
Somebody has sent you Korean text and you need to form a view about it before you pass it on. That is the situation this page is built for. Paste the text, run the check, and you get back a written description of what the writing looks like: which patterns are present, which are not, and where they cluster.
What you do not get is a number to forward to someone else. That is deliberate. A percentage travels through an organisation as a fact and arrives as an accusation. A description of signals stays a description, which is all anyone is entitled to draw from a piece of finished text.
What to look for
The Korean patterns worth checking in a draft
You can read for most of these yourself once you know they exist, which is the point of listing them. Korean-language detector pages elsewhere skip this entirely and talk about perplexity instead, which tells a reviewer nothing actionable.
- Register holds or slips
- Scan the verb endings down the page. One document should sit in one speech level. A slide from 합니다 into 해요 and back inside a section is worth a second look.
- Spacing accuracy
- Korean 띄어쓰기 is difficult enough that fluent writers get it wrong in ordinary prose. A long casual passage with no spacing slips at all is an odd thing to find.
- Loanword forms
- Check a handful of borrowed words. Human drafts mix 메세지 with 메시지. A draft that only ever uses the standard forms has been edited by someone or something.
- Particle rhythm
- 은/는 and 이/가 are a writer's emphasis controls. Paragraphs where every sentence takes the identical frame are paragraphs where nobody chose.
- Sentence endings
- Count how many sentences close on ~것입니다 or ~수 있습니다 in a row. Human paragraphs break the run with something short.
- Abstraction load
- Stacked ~적 and ~화 forms where a plain verb would do. Common in translated and generated text alike.
The three steps
How to run the Korean AI detector on a draft in three steps
Budget the runs as you go, because you get five of them in a rolling day.
- Step 1
- Paste the Korean text, up to 12,000 characters. For a long document, take the section you actually doubt rather than the whole thing.
- Step 2
- Run the check. The text is sent to our server and the description comes back.
- Step 3
- Read it next to the draft, and treat anything it raises as a question to ask the writer, not an answer about them.
Working within the limits
Five checks a day, and how to spend them
The allowance is five runs per rolling 24 hours, counted on our server, and the input cap is 12,000 characters. Rolling means there is no midnight reset to wait for: a run drops off your count 24 hours after you made it.
Twelve thousand characters is more Korean than it sounds, because one syllable block is one character. Most single essays and 리포트 fit in one run. If you are reviewing a stack of submissions, check the passage you have a question about rather than pasting each document whole, and you will get through far more than five documents on five runs.
Straight answer
Can this Korean AI detector name the model?
No. The instruction our server sends with every run tells the engine never to claim certainty about authorship or about which AI model wrote a text, and to describe observable signals and limitations instead. That constraint is in the request itself, not a disclaimer added underneath the result.
So if your workflow needs the sentence "this was written by ChatGPT", no tool can honestly supply it, including the ones that advertise it. What a review process can use is the description, plus whatever the writer can show you about how the draft came to exist.
The failure mode
Why good Korean writers get flagged
The writing most likely to trip a detector is the writing your organisation asked for. A 보고서 filled to a template, a 자기소개서 polished through twenty passes, an abstract written in careful 합쇼체: all of them are deliberately regular, and regularity is what detectors read as machine-like.
The bias is documented outside Korean too. "GPT detectors are biased against non-native English writers", by Liang, Yuksekgonul, Mao, Wu and Zou, reports that the detectors tested consistently misclassified non-native English writing as AI-generated while classifying native writing correctly, and suggests these detectors may penalise writers with constrained linguistic expression. If you review English work from Korean writers, that result belongs in your process.
If it goes wrong
What to do when a Korean draft is wrongly flagged
Whether you are the writer or the reviewer, the useful move is the same: stop arguing about the score and go looking for the writing process.
- Ask for the history
- 한글, Word and Google Docs all keep version history. A document that grew over days looks nothing like one that arrived complete.
- Ask for the working material
- Notes, sources, a discarded opening. Real drafting leaves debris.
- Ask which signal fired
- If a third-party tool produced the flag, ask what it actually reported. A bare percentage is not something a person can answer.
- Do not ask for a rewrite
- Telling a writer to make their Korean look less tidy degrades the work and settles nothing.
Who it's for
Who reaches for a Korean AI detector, and when
Editors handed Korean copy by a freelancer they have not worked with before. Teaching staff reading submitted work. Translation reviewers checking a Korean target text against what the source justified. Recruiters reading a 자기소개서 that reads a little too evenly.
And, more often than any of those, writers checking their own Korean before they send it, to see what a reader with a suspicious eye might pick up on. That use is the one this page is most confident about, because the person running the check already knows the answer and is only testing the appearance.