Signals are not proof
A writing checker may look at sentence-length variation, repeated words, connector patterns, punctuation rhythm, and common template phrases. These are observable features of a text, not evidence of who wrote it. A careful human writer can produce regular prose, and an AI-assisted draft can contain personal, irregular language. The same paragraph can receive different results after translation, editing, or a change of detector.
What a writing-signal check can observe
- Sentence-length consistency. Sentences that cluster unusually closely together can be worth reading again, but a form with a strict style can create the same pattern.
- Word repetition and vocabulary variety. Repeating a key term may be correct in a technical report; low repetition is not proof of human authorship.
- Connectors and template phrases. Frequent transitions such as “therefore” or “in conclusion” can make prose feel formulaic, especially when every paragraph follows the same structure.
- Repeated sentence openings. Several paragraphs that start with the same construction may need a rhythm edit, regardless of how the draft was produced.
- Punctuation rhythm and paragraph length. Similar comma patterns or equally sized paragraphs are editing clues, not identity markers.
Why scores change
- Short samples are unstable. Document Forge treats a sample under 180 characters or three sentences as too short for a stable comparison. A longer sample of at least 600 characters and eight sentences gives the browser more variation to describe, but it is still not proof.
- Language matters. Korean, English, and other languages have different sentence boundaries, spacing rules, and common connective phrases. A detector calibrated on one language should not be assumed to have the same meaning in another.
- Editing changes the signal. Rewriting, translating, combining sources, or adding quotations can change the pattern without changing the origin.
- Different tools use different references. A score from one detector cannot be compared directly with a score from another. There is no universal percentage scale.
A responsible way to use a checker
- Use a substantial sample that you are allowed to analyze, and remove private information that is not needed for editing.
- Read the explanation behind the result instead of treating the number as a verdict.
- Use the strongest signals as revision prompts: vary a repeated opening, replace an unnecessary transition, or split a sentence that carries too many ideas.
- Compare the output with your own drafts, notes, sources, and revision history. Those materials say more about authorship than a pattern score.
- Ask for human review when the result could affect a grade, job, admission, or accusation.
Do not revise only to defeat a detector
Replacing words with a thesaurus, adding random mistakes, or forcing unnatural sentence lengths can make a document less clear while doing nothing to establish who wrote it. Edit for a reader: use concrete nouns, remove padding, put the important qualification where it belongs, and make each paragraph earn its place. A good revision should still be good when no detector is present.
Document Forge's deliberately narrow result
The browser analyzer reports an interpretable reference score based on the signals described above. It labels confidence based on sample length and lists contributing signals so you can inspect the reason for the result. It does not identify authorship, detect plagiarism, or guarantee that a text was or was not produced by AI. Text stays in the browser for this default analysis; an optional server model is a separate, consent-gated path and should not receive confidential text.
Questions to answer before relying on a result
- How long was the sample, and is it long enough for the tool's stated boundary?
- Which language, genre, quotations, and formatting rules shaped the prose?
- Does the result agree with the draft history and source notes?
- Would a reasonable human reviewer reach the same conclusion without the score?