How does this AI text detector work?
The detector is an algorithm we wrote for Zyphh. It runs on your device and measures six things: how much sentence length varies, how often sentences open the same way, repeated four-word sequences, stock phrases, how concentrated the vocabulary is, and how many sentences lean on transition openers. It doesn't call a paid detection service or a language model. The rules and weights are laid out on the methodology page.
The result is an experimental style index. It helps you see why a passage reads as formulaic, and it can't tell you who wrote it. Nobody has calibrated the rules or weights against a labeled corpus, so human writing can score high and generated writing can score low. Even when the tool finds several AI-like patterns, authorship stays unknown.
How do I check a passage?
Paste English prose or open a UTF-8 text file. A document-level index needs 80 to 3,000 words, at least five sentences and no more than 20,000 characters. Shorter samples, other languages and code-heavy text get an inconclusive result along with the reason. The language check is a simple heuristic, so specialist English is occasionally excluded by mistake.
Use the original, unedited text. You'll see each feature's contribution, the underlying statistics, exact highlights for repeated phrases and openings, and a sentence-length map. You can export the full report, which includes the engine version and your original text. Nothing gets published to a public URL.
What does the 0 to 100 score mean?
The index is the sum of six hand-written rules. It isn't the probability that AI wrote the text, the share of AI-written words, or a calibrated confidence score. More points mean more of this algorithm's chosen patterns showed up.
Below 35, the text shows few AI-like patterns. From 35 up, the signals are mixed. A score of 65 or higher with three contributing features counts as more AI-like. These cutoffs are exploratory product rules, and no accuracy study stands behind them. A low score doesn't certify human authorship, and a high score doesn't establish misconduct. For the reasoning behind that caution, read how accurate AI detectors are.
What's the difference between detection, watermarking and rewriting?
Style analysis describes the writing in front of it. Watermark verification is a separate job: it checks for a signal deliberately embedded during generation, and it needs a compatible method and configuration. This detector doesn't verify watermarks, and it can't tell ChatGPT from Claude or Gemini or name a model release.
For hidden characters, use the watermark checker, which maps unusual Unicode. For wording, the paraphraser suggests edits from curated phrase and synonym rules. If a style score moves after you edit, you've learned that the score is sensitive to those edits. You haven't learned that the original was AI-written or that a watermark is gone.
Is my text private?
For the detector and the rewriter, yes. Both run in a browser worker, so your text never goes to a detection vendor, an AI service or our server. You don't need an API key or an account. Page assets and public configuration download as usual, but those requests carry nothing you submitted.
The text stays in page memory. The app doesn't write it to localStorage or a database, and exported files stay on your device. When ads are enabled, they appear only on educational guide pages, never near the tools or results. The privacy page covers hosting metadata and the optional watermark-verifier integration, which is controlled separately.