What the detector is actually measuring
This tool doesn’t “know” whether AI wrote your text — no detector does. It measures statistical fingerprints that tend to differ between human and machine writing, then estimates a probability:
- Sentence uniformity — AI tends to produce sentences of similar length; humans mix short punchy lines with long winding ones (this variance is called burstiness).
- Vocabulary richness — the ratio of unique words to total words; machine text often leans on common, high-probability words.
- Predictability — text a language model finds unsurprising (low perplexity) is more likely machine-generated, because models produce what their own distributions predict well.
- Structural tells — overuse of transition words (“furthermore,” “moreover”), uniform paragraph sizes, repetitive sentence openings.
None of these is decisive. They’re correlations, and correlations produce both false positives and false negatives.
Why the arms race makes detection unreliable
The detector and the generator are playing the same game with the same rulebook. Anyone can lower an AI-probability score by paraphrasing, varying sentence length, or mixing in human edits — the very signals the detector reads are the ones a motivated writer can flatten. Meanwhile, formal human writing (academic abstracts, technical docs, legal boilerplate) naturally exhibits the “AI-like” uniformity that trips the tool. The result is a tool that’s easiest to fool exactly when it matters most and most likely to misfire on careful human prose.
How to read the score responsibly
| Score band | Reasonable reading |
|---|---|
| High AI probability | ”Worth a closer human look” — not “this is AI” |
| ~50% | Mixed, formal, or too short to judge — inconclusive |
| Low AI probability | Bursty, varied prose — but editing can produce this from AI too |
Longer samples (200+ words) give more stable signals than short snippets, but more data doesn’t fix the underlying false-positive problem.
A better posture than detection
If your real goal is academic or editorial integrity, process beats forensics: assess drafts and revision history, use in-person or oral components, and design assignments that AI can’t trivially complete. Detection scores can inform a conversation — they should never replace one.