About
How It Works
LexyAI analyses ten independent properties of any text and combines them into a single verdict. Everything runs inside your browser — nothing is sent anywhere. Here is what we actually look at, and why it matters.
01
Complexity Shift
Human writing changes rhythm and density as the author's thinking shifts — a burst of detail, then a simpler observation, then a complex argument. AI keeps its complexity unnaturally constant from start to finish, because it is always optimising for coherence.
What it catches: text that reads at the same "depth" throughout, with no intellectual peaks or valleys.
02
Rhythm Variation
Humans write short sentences. And then occasionally they write very long ones that wind through multiple ideas, connecting things in ways that feel unplanned. AI prefers sentences of similar, medium length — it avoids extremes because extremes hurt coherence scores.
What it catches: text where every sentence is roughly the same length, with no short punches or long runs.
03
Word Distribution
In any natural language, a few words appear constantly and many appear rarely. AI leans too hard on the common ones and avoids the rare ones — the result is a word-frequency distribution that fits the expected mathematical curve almost perfectly. Human writing is messier, more personal, more unpredictable.
What it catches: vocabulary that is too statistically "safe" — no unusual choices, no personal vocabulary, no surprises.
04
Vocabulary Range
How often does the text introduce a word it has not used before? AI recycles the same pool of reliable, common words. Humans reach for variety — different ways to say the same thing, more specific terms, less expected choices.
What it catches: narrow, repetitive vocabulary that recycles the same words across the whole text.
05
Phrase Variety
AI reuses the same two-word constructions repeatedly — transitional phrases, connectors, openers. Humans naturally vary how they link ideas and rarely repeat the exact same pairing more than once.
What it catches: repeated phrase patterns that suggest the text was assembled from statistical building blocks.
06
Text Predictability
AI-generated text is built from repeating statistical patterns — which means it compresses unusually well. Think of it like a ZIP file: the more repetitive the underlying structure, the smaller it gets. Human writing is noisier, more individual, and resists compression.
What it catches: text that is structurally too regular to have come from an unpredictable human mind.
07
Vocabulary Growth
As a human writes more, new words enter at irregular intervals — a new topic suddenly brings a cluster of fresh vocabulary. AI introduces new words at a smooth, predictable rate from start to finish. The growth curve for AI is a clean mathematical line. For humans, it jumps.
What it catches: vocabulary that grows too smoothly, without the bursts you get when a human shifts topic or emphasis.
08
Generic Phrasing
AI has a set of favourite phrases it reaches for constantly — "in order to", "it is worth noting", "plays a crucial role", "first and foremost", and around sixty others. These are not wrong, they are just overused by AI because they are statistically reliable. Humans tend to phrase things more personally.
What it catches: text dense with stock phrases that no individual human would repeat so frequently.
09
Short Fragments
Humans write incomplete sentences. "Right?" "No." "Exactly." These micro-fragments — sentences of three words or fewer — appear naturally in human writing and almost never in AI output. AI is trained to be grammatically complete at every step.
What it catches: absence of any short fragments, which is statistically unusual in genuine human writing.
10
Punctuation Rhythm
Humans place commas, dashes, and colons erratically — sometimes three in quick succession, sometimes none for a long stretch. AI distributes punctuation at eerily regular intervals, because it is predicting the next token one step at a time with no sense of overall rhythm.
What it catches: punctuation that is spaced too evenly across the text.
Verdict
How the Score Is Calculated
Each signal produces a score from 0 to 100. The ten scores are combined with different weights — signals that are empirically more reliable carry more influence. The final composite score maps to four verdicts:
| Score | Verdict | What it means |
|---|---|---|
| 72 – 100 | Likely AI | Multiple signals consistently point to machine authorship. |
| 52 – 71 | Possibly AI | Mixed signals — may be AI-assisted, lightly edited, or heavily structured. |
| 32 – 51 | Likely Human | Most signals lean toward human authorship. |
| 0 – 31 | Human | Strong evidence of genuine human authorship across all signals. |
Limits
What LexyAI Cannot Do
LexyAI is a statistical tool, not a definitive classifier. It has real limits you should know about:
- Texts under 50 words are too short for reliable results.
- AI text that has been rewritten or heavily edited by a human will score lower.
- Formal human writing — academic papers, legal documents, official reports — can statistically resemble AI output and may score higher than expected.
- Results are indicators, not proof. Do not use them as sole evidence in academic, legal, or professional decisions.
All analysis runs inside your browser. Nothing is ever sent to a server.