AI detector
AI 탐지기
한국어로 읽기An AI detector estimates from statistical features how likely text is AI-generated; some contests require a report, but false positives are common.
Also written as: AI 판별기 · AI 검출 · AI 텍스트 탐지 · AI detection · GPTZero · AI 탐지 보고서
Meaning
An AI detector is a tool that statistically estimates the probability that text was generated by AI. It measures features such as predictability (perplexity), variation in sentence length (burstiness) and vocabulary distribution and reports "N% likely AI." It reads patterns rather than watermarks, so it is not a definitive verdict. Uniform, smooth human prose gets flagged as AI, and revised AI prose can pass as human. Korean-language detectors in particular have less training data than English ones and vary widely in accuracy.
Why it matters
Detectors became important because they turned into a submission requirement. As literary prizes and some contests banned AI use, publishers and organizers began asking entrants to attach a detection report. For writers a detector means two things. First, a reference for how "AI-like" a manuscript reads. If the score drops after a revision pass to remove tells, the revision worked. Second, the risk of false positives. Cases of entirely hand-written work being flagged and disqualified have been reported, with little recourse. So a detector score should be used as an inspection tool, not a pass mark. Open the paragraphs that score high and you usually find AI-tell patterns (uniform sentences, emotion nouns, summarizing endings), and fixing those improves the prose for human readers too. Gaming the score (deliberate typos, scrambled sentences) ruins the writing.
Example
Before submitting to a contest, Seojun ran his manuscript through a detector. Overall: 34% AI. By paragraph, three were above 80%.
He opened them. One explained the hunter association's rank system: seven sentences, all ending the same way, all similar in length. The other two were the protagonist's reflections after a fight. "Relief washed over him." "He steeled a new resolve."
He turned the explanation into dialogue and the reflections into actions (cleaning gear, checking the settlement statement). Rerun: 12% overall. But what he had gained was not the score; it was a map of which paragraphs were weak.
Common mistakes
- Treating the score as a verdict. A detector estimates. Human writing can score 40%.
- Damaging the prose to lower the score. Typos, broken grammar and scrambling may fool the detector but not the reader.
- Substituting a detector score for the contest rules. Rules ask whether AI was used. A low score does not make a violation not a violation. Read the rules first.