ai-novel-audit

Run the audit workflow on the text: execute the local signal scripts for quantitative metrics, then score six qualitative dimensions with quoted evidence. The output is an AI-authorship percentage range with a confidence level, not a definitive verdict.

Key signals include uniform sentence length with low burstiness CV, high cliche density across multiple categories, named emotions instead of shown ones, generic interchangeable details, stock metaphors, and paragraph-level verbatim repetition. No single signal is conclusive on its own.

No. All AI text detection is probabilistic and can misjudge non-native writers, formulaic genre fiction, and heavily edited AI drafts. High-stakes decisions require provenance evidence such as draft history and version records alongside stylistic analysis.

It runs, but the report must state low confidence because short texts produce noisy signals. The Skill explicitly refuses to give high-confidence scores when the sample is too short or signals contradict each other.

Romance and fantasy web novels naturally use formulaic phrases, which trigger cliche-density signals. The audit compensates by weighing detail specificity and voice consistency more heavily to distinguish formulaic human writing from AI output.

Local scripts approximate external detectors using proxies like entropy, TTR, and bigram surprise curves without a real language model perplexity score. Results are risk signals for editorial review, not author-identity verdicts, and external platform spot-checks are recorded separately.