How this AI text detector works
A transparent, browser-only heuristic. No model API. No claim of commercial detector accuracy.
What we score
When you click Analyze, JavaScript on this page measures a few stylistic signals that often appear in raw large-language-model drafts. The signals are blended into a single 0–100 “AI-like” score for exploration only.
- Sentence-length burstiness. Human writing usually varies sentence length. Very even rhythm across many sentences pushes the score up.
- Vocabulary variety (type-token ratio). Repeated word choices and low variety relative to length look more template-like.
- Stock LLM phrases. Patterns such as “delve into”, “it’s important to note”, or “in today’s digital landscape” add weight. The list is illustrative, not exhaustive.
- Formal transition density. Heavy use of “furthermore”, “moreover”, “in conclusion”, and similar glue words is common in unedited model output.
- Punctuation uniformity. Extremely tidy, uniform punctuation with little stylistic noise can nudge the score.
What stays on your device
Scoring runs in your browser. The pasted text is not sent to our servers for analysis. Hosting logs (if Cloudflare Pages is used later) may see normal page requests, not your textarea content. See Privacy for the full wording.
What this is not
This is not GPTZero, Originality.ai, Copyleaks, Turnitin, or any trained classifier. Those products use labeled datasets and proprietary models. Our score is a drafting aid: useful for a quick vibe check, useless as academic, HR, or legal proof.
English only (for now)
Phrase lists and rhythm assumptions are tuned for English. Other languages will produce unreliable numbers. We would rather say that than pretend multilingual accuracy.
How to use the score
Treat a high score as “this draft looks very smooth and template-like,” not “a machine wrote this.” Edit for voice, concrete detail, and uneven rhythm if you want a more human feel. For high-stakes decisions, use a dedicated commercial detector plus human review.