“95% AI-generated,” said the detector about a paragraph I typed myself — slowly, grumpily, at midnight. Terse human prose scores as artificial because detectors measure statistical regularity, and short texts have no room to be irregular. Meanwhile obviously-spun articles sail through at 12% because length buys variance. This guide explains burstiness and n-gram overlap honestly, sets the 400-word reliability floor, and shows checking workflows that avoid false accusations — including why our detector shows evidence instead of verdicts.
Part of the SEO publishing guide. Check with evidence in the AI detector; verify originality in the plagiarism checker; mark up only visible copy with the FAQ schema generator.
What detectors actually measure (burstiness + 5-grams)
| Signal | Human pattern | AI pattern | Failure mode |
|---|---|---|---|
| Burstiness | Varied sentence lengths (4–28 words) | Uniform ~18-word sentences | Short texts can't vary |
| Perplexity | Surprising word choices | Predictable phrasing | Technical writing looks “predictable” |
| 5-gram overlap | Unique phrasing | Training-data echoes | Quotes and boilerplate false-flag |
Our detector surfaces these three numbers with a sliding 5-gram window view instead of a single percentage — because a 72% verdict on 400 words with visible uniform-length runs means something, while 95% on 50 words means nothing. Independent October 2026 coverage keeps confirming vendor accuracy claims overreach; tools that hide methodology deserve the least trust, not the most.
The 400-word floor and the rewrite protocol
- Under ~150 words: do not test — detectors misfire on terse human prose routinely. No verdict is valid here.
- 150–400 words: indicative only; corroborate with process evidence (drafts, edit history) before any conclusion.
- 400+ words: patterns stabilize; burstiness variance plus n-gram runs become meaningful — still evidence, not proof.
- On flags: rewrite passages for variance (sentence lengths, transitions, concrete detail) rather than re-rolling through paraphrasers — adversarial laundering degrades quality while teaching nothing.
Editors: pair detection with the plagiarism checker (originality is the checkable claim) and publish the workflow, not just the verdict. A 96% uniqueness score with visible burstiness beats any single detector number.
Never gate grades, jobs or accounts on output
The strongest statement in this guide: detector output must not decide academic, hiring or moderation outcomes alone. False-positive rates on non-native writing run multiples higher (simpler constructions pattern-match “AI-like”), creating discrimination risk stacked on accuracy risk. Institutions need process policies (drafts, vivas, revision trails); platforms need human review; individuals accused deserve the evidence, not a percentage. Our tool exists to inform revision — which passages read uniformly and how to vary them — not to certify authorship. Anyone selling certainty here is selling something else.
General guidance only, not academic-integrity or legal advice. Detection is statistical evidence with known failure modes — treat it accordingly.