How It Works

01 / Complexity Shift

Does the text breathe?

LexyAI measures Shannon entropy across moving windows of the passage. The useful quantity is not entropy alone, but how much information density changes as the text progresses.

A very flat local profile can contribute to an AI-associated score, while larger shifts can contribute to a human-associated profile. Genre still matters, so this signal is interpreted with the others.

02 / Rhythm Variation

How much do sentence lengths move?

Sentence burstiness uses the coefficient of variation of sentence lengths. Passages that repeatedly use medium-length sentences have a different rhythm from writing that mixes short fragments, long explanations and abrupt changes of pace.

This is a structural clue rather than an authorship test. Academic, legal and technical writing can naturally be uniform.

03 / Word Distribution

Does vocabulary follow a clean frequency curve?

LexyAI calculates a Zipf-style fit over ranked word frequencies. Language naturally follows heavy-tailed distributions, so a strong fit is not itself evidence of AI.

The value comes from combining word-frequency structure with lexical diversity, repetition and other measurements.

04 / Vocabulary Range

How broad is the local vocabulary?

Moving-window lexical diversity measures how widely a passage uses vocabulary across local sections instead of relying only on a total unique-word count.

This helps distinguish a text that is consistently narrow from one that introduces new terminology as topics change.

05 / Phrase Variety

Does the text repeat short constructions?

Repeated adjacent word pairs can reveal formulaic phrasing, habitual transitions or topic-specific terminology. LexyAI measures repeated bigram rate relative to all available adjacent pairs.

Common phrases are not inherently machine-written, so repetition is treated as a supporting signal rather than a verdict.

06 / Text Predictability

How easily does the text compress?

A lightweight LZW compression measurement provides a proxy for character-level regularity. Repeated structures can be represented more efficiently than highly irregular text.

Compression is affected by language, formatting and repetition. It is therefore one feature in the overall statistical profile.

07 / Vocabulary Growth

How smoothly are new words introduced?

LexyAI samples cumulative type-token ratio as the text grows and examines how closely that curve follows a smooth decay pattern.

Irregular growth can occur when a writer changes subject, emphasis or vocabulary abruptly; generated passages can sometimes exhibit smoother progression.

08 / Generic Phrasing

Does the passage rely on stock connectors?

A curated list of common formal connectors and constructions is used to measure density. These expressions are normal English, so the detector looks at concentration rather than treating any single phrase as an AI marker.

Connector density is deliberately kept from dominating the composite score.

09 / Short Fragments

Are there micro-sentences?

Very short fragments such as “No.” or “Exactly.” change sentence rhythm and can be common in conversational prose. LexyAI measures their proportion among detected sentences.

Formal prose naturally uses fewer fragments, so their absence is not proof of machine writing.

10 / Punctuation Rhythm

How evenly is punctuation distributed?

LexyAI examines the spacing of selected punctuation marks and normalizes the result with the number of observations available.

The measure describes one aspect of punctuation rhythm and is interpreted together with sentence and lexical features.

Composite / Score

How are the signals combined?

Each signal is mapped to a bounded 0–100 contribution and combined using fixed weights. Structural signals carry more influence than simple phrase-list features. The engine then calculates signal agreement and applies reliability adjustments based on the amount of text structure available.

The headline score is a pattern score, not a measured probability that a machine wrote the passage. A separate confidence value describes how consistently the signals agree and whether enough observations are available for a stable estimate.

Limits / Important

What LexyAI cannot prove

• It cannot prove that a specific person or AI model wrote a passage.
• Human writing can produce AI-associated statistical patterns.
• Edited or paraphrased AI text can lose detectable regularities.
• Short or highly specialized text provides fewer useful observations.
• Results should not be treated as sole evidence in academic, legal, employment or disciplinary decisions.

For the strongest interpretation, inspect the complete signal profile rather than relying on a single percentage.