검사는 이렇게 진행됩니다

붙여넣는 글

동기들에게 돌린 설문에 47명이 답했습니다. 대표성 있는 표본을 고르는 것이 신뢰할 수 있는 결론에 핵심입니다. 그래서 학과 게시판에도 올렸습니다.

82 글자

돌려받는 결과

예시

34%AI처럼 보임

수정 필요

GPTZero: 41%
Turnitin: 36%
Originality: 25%

Why AI detection matters

AI writing tools created a crisis of authenticity in academia. ChatGPT is useful for brainstorming. Submitting AI-generated work as your own is not. Before you turn a paper in, see which sentences look machine-made.

What you do here

Line by line: not one percentage — the shaky sentences.

Rewrite here: the paraphraser works on the flagged lines.

The text stays yours: we do not store or train on what you paste.

The science: perplexity and burstiness

The detector looks for patterns typical of large language models. Two measures do most of the work.

Perplexity

How “surprised” a language model is by the next word. AI text often has low perplexity — smooth, generic. Human writing is full of unexpected choices.

Burstiness

Variation in sentence length and structure. Humans mix short and long. Models tend to stay even.

AI vs human writing

The difference is often subtle. The model looks at structure, vocabulary, and tone — not a watermark.

Sentence structure. Human: varied, sometimes fragmented. AI: grammatically clean, often repetitive.

Vocabulary. Human: specific, niche. AI: high-frequency, generic.

Tone. Human: uneven, subjective. AI: polite, even, objective.

Frequently asked questions

No detector is perfect. Use this as a diagnostic, not a verdict.

Can AI detectors be wrong?

Yes. False positives happen, especially on formal academic prose. That is why we mark sentences, not just a score.

Does paraphrasing bypass detection?

Simple synonym swaps usually do not. Heavily rewriting and adding your own thinking will lower the score — because you added human work.

Is this safe for confidential documents?

We do not store, share, or train on the text you submit.

AI 탐지기