Transparency at Vairify
Methodology & limitations
Last updated: October 8, 2026
What the Human Crafted Confidence estimate is
Version 3.0 splits each page into sections and estimates, for each section with at least 40 words and 3 sentences of prose, how strongly its style resembles machine written text. The page’s Human Crafted Confidence is the word-weighted average across scored sections. Sections are labeled “Likely machine written patterns” at a machine-pattern score of 0.70 or higher, “Mixed authorship signals” from 0.50 to 0.70, and “Likely human crafted patterns” below 0.50. These are probabilistic estimates of writing style, not proof of who or what wrote a page.
Only prose is estimated: text ending in sentence punctuation, or descriptive lines of eight or more words. Short headings, labels, names and menu items are left out.
What the estimate looks at
- Personal texture: contractions, questions, quotations, parenthetical asides, first-person voice, citations and specific figures. Machine written text tends to have little of this.
- Hedging: words such as relatively, tends, appears and may.
- Formulaic structure: three-part lists and “not just X” contrasts.
- Stock vocabulary: words that AI writing tools overuse, such as seamless, crucial and transformative.
- Repeated sentence openings: many sentences starting with the same word.
Testing and uncertainty
We do not publish a validated accuracy rate for this version. The estimate has not been independently evaluated across the kinds of pages people may submit. Short sections, heavily edited AI drafts, formal or templated human writing, and writing by people with different language backgrounds can all be misjudged. Treat every label as a prompt to inspect the text and its context.
Extraction and sample limits
We retrieve at most 2 MB of public HTML, follow at most four redirects and analyze at most 100,000 characters. We favor the first main or article element, remove common navigation and non-content elements, and exclude exact repeated blocks and nested duplicate extraction. Scripts are not executed, so dynamic content may be missing. CSS visibility is not fully evaluated. The extracted text is displayed so you can inspect coverage.
Page-level writing statistics require at least 200 word tokens and eight sentences with three or more words; the Human Crafted Confidence estimate requires at least one scored section and 80 scored words. These are conservative product thresholds, not validated accuracy guarantees. Tokenization recognizes Unicode letters and internal apostrophes; sentence segmentation uses English conventions. Only English text is scored. Language is judged from the page’s text, because declared language tags are often missing or wrong; pages with very little text fall back to the declared tag.
How to read each metric
- Sentence-length variation: population standard deviation divided by mean sentence length, multiplied by 100. This coefficient of variation can exceed 100%.
- Repeated four-word phrases: repeated occurrences beyond the first occurrence divided by all four-word windows, multiplied by 100.
- Vocabulary diversity: average distinct word count across complete, non-overlapping 100-word windows; a trailing incomplete window is excluded.
Limits of the estimate
The estimate has not been independently evaluated, and we do not claim high detection accuracy. A stronger evaluation would need larger provenance-labeled samples across domains, AI tools, edits, lengths and language backgrounds, with published false-positive and false-negative rates.