What you get

ai_detect_en.py, the detector (pure standard library, Python 3.8+, zero dependencies)
example_run.py, runnable example, prints a side-by-side result
lm_en.bin + model_params.json, trained model data (~88 KB)
scorer.py, optional local web page (paste text in a browser)
Runs fully offline, no API key, no signup, nothing you paste leaves your machine

Indicative accuracy (held-out test set; your results will vary by model, length and genre)

Metric	strict	balanced
False positives on real human text (5035 samples, 12 sources; 0.6% source-balanced)	0.9%	5.0%
Detection of AI-generated text (held-out calibration set)	~91%	~98%
Detection on other model families	varies*	,
False positives on modern / genre fiction (1900–1930 novels)	~1%	,
Detection of AI-written fiction	79%	94%

* Other-model-family figures are from early, partial tests only and are not independently verified. Major commercial models are not yet in the test set; treat every score as indicative, never definitive.

Human corpus: Project Gutenberg (19th-century novels + 138 public-domain 1900–1930 genre novels), pulp fiction, Wikipedia, Stack Exchange, Hacker News, Guardian news, real published text.

Pricing, one-time offline licence, with optional yearly updates

Tier	Price	License
Free	$0	The web checker on this site, 5 checks/day, no signup (not the offline pack)
Personal	$19 one-time	One individual, unlimited runs
Commercial	$99 one-time	Also permits embedding the detector in one of your own products, up to 5M words/year, see licence
Commercial+	$299/year	Same embedding rights as Commercial, but no annual volume cap
OEM	from $999/year	Multiple products, white-labelling, no volume cap, ask us

Comparable detectors are subscription-based and metered, GPTZero from $12.99/mo, Originality.ai $14.95/mo, Winston AI $18/mo, Pangram $20/mo. This pack is a one-time purchase that runs offline with no word limits: roughly one month of a competitor's fee, for permanent use.

Your genre, language or source material not represented in the numbers above? We offer a threshold calibration service, $299, one-off: send us a sample of the text you actually work with and we tune the detector's thresholds against it, returning a calibrated build plus a short report. Order the calibration service →

What this is NOT, please read before buying

It is not proof of authorship. It is a statistical estimate, do not use it as sole evidence in any disciplinary, legal, or editorial decision.
It is not a "humanizer", not an "undetectable" tool, and not a service to bypass AI detectors. We sell detection, not evasion.
Known limitations: Major commercial models not yet independently tested. On other model families, early partial tests suggest detection is possible but results vary widely and are not yet verified, treat every score as indicative, never definitive. News-style text still scores higher (7.7% false positives, source-balanced); non-English text, lists, tables, résumés, job posts and code are returned as not_applicable.
Academic integrity: this tool must not be used to misrepresent the authorship of academic work, to complete assessments for someone else, or to circumvent your institution's rules.

Built on a language model trained on real human English corpora, combined with 18 measurable text statistics, sentence and paragraph-length burstiness, lexical diversity, punctuation distribution, function-word deviation, and bigram repetition. No banned-word list, no black box.

Personal tier: one individual, unlimited runs. Commercial tier ($99): also permits embedding the detector in one product or service of your own, up to 5 million words per year, but not offering detection itself as a service to others, and not white-labelling. A wider scope needs Commercial+ ($299/year, same embedding rights, but no annual volume cap) or an OEM licence, from $999/year (multiple products, white-labelling, no volume cap), ask about OEM at support@prosechecker.com. Resale of the pack itself is not permitted under any tier. Updates: 12 months included, optional after that. The free online checker is limited to 5 checks per day, the pack has no limits at all. Estimate only, not proof of authorship, and not licensed for academic-integrity circumvention.

New on the blog: Why AI detectors keep accusing real people , books written before ChatGPT existed come back as “29.7% AI”. Here is why, and what a false-positive rate actually tells you.

Get weekly AI-detection insights

One practical email per week on detecting AI-written text, what changed, what to watch, how to read a score. No spam; unsubscribe anytime.

We use your address only to send this weekly email. One-click unsubscribe in every email.

AI-Text Detector