Back to analysis

Methods and tests

The version 1.0.0 benchmark includes 24 texts. In each language, one of six AI-generated texts reached the test threshold. The AI style signal supports editorial review of writing patterns, while authorship remains undetermined.

LanguageHuman / AI source textsHuman / AI scoredHuman recallAI recallFalse positivesAvg. AI / human
DE6 / 66 / 66 / 61 / 60 %35.5 / 28.3
EN6 / 64 / 64 / 41 / 60 %36 / 22.3

Benchmark design

The same analyzer processed six passages published before the LLM era and six ChatGPT texts per language. Human samples include fiction, letters, a parable, and argumentative prose. Five AI texts per language use a natural style and one is a deliberate stock-phrase stress case. The small sample has limited coverage of modern web content, models, and mixed authorship.

Threshold and coverage

The threshold was fixed at 60 before the run. An AI score of 60 or above counts as a true positive. A human score at or above 60 counts as a false positive. Two English passages contained fewer than five complete sentences and received no percentage. The table shows false positives among scored texts only.

Where the style signal misses

The five more naturally worded AI texts in each language scored below the threshold. Both deliberate stock-phrase cases scored 92. Repeated phrases, stock phrasing, similar sentence lengths, and transitions affect the score. Genres and editorial style guides can produce similar patterns. The percentage scale is uncalibrated.

Language, SEO and GEO measures

The analyzer approximates syllables for the German and English Flesch measures and flags specific grammar and punctuation patterns. SEO covers topic fit, search intent, structure, and available metadata. GEO covers direct answers, structure, evidence signals, examples, and scope. Links, evidence quality, and actual citations require independent review.

Functional verification

Checks covered text analysis, language limits, exact highlights, numerical facts during revision, PDF structure, HTML extraction, DNS and redirects, size limits, and SEO metadata. Browser checks covered highlights, revision, PDF, and the language toggle. The mobile layout was inspected at a width of 390 pixels.

Import and revision limits

The importer works with publicly reachable HTTPS content and reads the response without running page JavaScript. Access restrictions, large pages and sign-in walls can prevent retrieval. The rule-based revision makes limited editorial changes and identifies places where examples or evidence would improve the text.

Other languages

The reported results cover the German and English samples. French, Spanish and Italian require separate language profiles and tests across genres.

Sources