How Humanizer AI detection works

Method description reviewed on 8 September 2026 against the application implementation. This is a description of the configured methods, not an independent accuracy benchmark or confirmation that every model is available on every request.

Start with the right interpretation

The web detector returns an AI-likeness score from 0 to 100. Higher values indicate more patterns associated with AI writing. If a view labels the result as a human score, it displays the complement: 100 minus the AI score. Neither number is a calibrated probability of authorship, a percentage of copied words, or an institutional Turnitin result.

For example, an illustrative AI score of 70 corresponds to a human score of 30. It does not establish that 70% of the words were generated by AI. This arithmetic example is not a recorded detector test.

A short-text example we observed

In a development-build check on 8 September 2026, this assistant-written, 20-word input returned 78 with “Likely AI”: “The library opens on Monday. Bring your card to borrow a book. Staff can help you find the reading room.” Below 50 words only the statistical rules run. This is one observed result, not a production benchmark or an accuracy estimate.

The web detector's three signals

• Statistical rules: writing patterns including sentence variation, readability, repetition and word distributions. These rules always run.
• GPT-2 perplexity: the Xenova GPT-2 model runs through Transformers.js on the server to measure how predictable the text is. Predictability is a signal, not proof of AI use.
• A fine-tuned DistilBERT classifier: the application attempts to load a classification model. Its score is combined with the rules and, when available, perplexity.

The weighting changes when the signals agree or disagree. If a model cannot load or an analysis fails, the available signals still produce a result; rules alone can be used. Sentence highlights may use the classifier or rules, so a highlight is not a word-level provenance finding. We do not present historical training comments as current real-world accuracy.

This web path does not request a Turnitin, Grammarly, GPTZero, Pangram or ZeroGPT score. The rewriting service is a separate workflow and may use different providers; its output is not validation by those detectors.

Length, language and access limits

The free web request limit is 1,000 words. Signed-in free accounts have a limit of three successful detections. The interface can ask an anonymous visitor to sign in; these checks are not an unlimited free service. Longer documents require a subscription. Humanization requires a paid plan.

Below 50 words, the web detector skips the GPT-2 and DistilBERT signals and uses rules. Very short text gives little evidence. Longer documents are split into chunks of up to 5,000 words. When there are more than three chunks, the first, middle and last are sampled; later chunks use a faster analysis without perplexity. A document-wide number can therefore summarize a sample, not every sentence.

The interface supports multiple languages, but that does not establish equal detection accuracy across them. We have not published a current, matched evaluation for every language or for Tagalog, mixed-language writing, poetry, code or specialist prose. Treat those results cautiously.

Web, API and extension are different

The public detection API currently uses the statistical detector, with API-key access for Pro and Business and a 10,000-word request limit. It does not run the same three-signal combination as the web detector. The extension has a separate detection route, cache and quota, and can combine rules with the classifier. Scores across these surfaces should not be treated as interchangeable. See the API documentation for integration details.

Privacy and saved work

Detection requires sending your text to our server. Text may be temporarily stored for processing under our privacy policy. Signed-in document features can save drafts and results, and some surfaces use local browser storage or caches. Do not assume every workflow is storage-free or confidential enough for sensitive submissions. Remove personal, confidential or unpublished third-party material unless you have permission and the relevant terms meet your requirements.

How we would measure accuracy

A useful evaluation needs known-provenance human and model-written passages, with permission to publish them. Freeze the corpus, language, length groups, generation prompts, model versions, date, detector version and decision threshold before testing. Keep edited and mixed-authorship passages in separate groups.

• Report human false positives as human passages flagged divided by all human passages.
• Report AI misses as generated passages not flagged divided by all generated passages.
• Publish counts, denominators, failures and uncertainty intervals alongside any overall accuracy.
• Review rewrites separately for preserved facts, quotations, meaning, grammar and naturalness. A lower detector score does not measure these qualities.
• Use a native reviewer before making language-specific quality claims.

This protocol is proposed evaluation guidance. We are not claiming that this full benchmark has been completed. Individual demonstrations and vendor reports are labeled separately in our reviews.

What to do with a disputed result

A false positive flags human writing; a false negative misses generated writing. Both are possible. Keep drafts, notes, sources and version history. Discuss the actual writing process and applicable policy instead of using a detector alone to accuse someone or decide a high-stakes outcome.

Open our AI detector · Understand Turnitin access · Grammarly review · GPTZero and ZeroGPT compared