Free AI-Text Detector

This is a statistical AI-text detector (offline N-gram language model + text statistics). Validated 2026-09-28 with strict train/test separation: real human text (5035 held-out samples, 12 sources, 19th-century novels, 138 public-domain 1900-1930 genre novels, pulp fiction, Wikipedia, Stack Exchange incl. ESL learners, Hacker News, news) → 0.9% false positives (strict) / 5.0% (balanced); 0.6% when each source is weighted equally. Real AI text (92 samples) → 91.3% detected (strict) / 97.8% (balanced). Fiction: 0.0% false positives on 19th-century literary fiction; ~1% on modern / pulp-style fiction (4231 windows), and AI-written fiction is now detected at 79% (strict) / 94% (balanced), up from 61% before genre fiction was added to the corpus. Guardrails: non-English text, lists, tables, code, and samples under ~80 words are returned as not_applicable instead of guessed. Limitations: major commercial models not yet tested; news-style text still scores higher (7.7% false positives, balanced across sources). An estimate only, not proof of authorship. Sold strictly as a detection tool: it is not a humaniser, it does not rewrite or “humanise” text, and it is not licensed for academic-integrity circumvention.