How the AI Detector Works
Our detector uses multiple research-backed algorithms running entirely in your browser. No text is ever sent to a server. Here's an overview of our approach.
Two Parallel Detection Engines
AI writing behaves differently depending on context. Academic essays, emails and reports contain different AI patterns than creative stories or social posts. Our detector runs two engines simultaneously and intelligently combines their results.
📄 Academic Engine
Optimized for essays, emails, reports and professional writing. Identifies the characteristic patterns that AI models produce when asked to explain, summarize or persuade.
✍️ Narrative Engine
Optimized for stories, creative writing and personal content. Detects the structural and stylistic patterns that emerge when AI generates long-form creative text.
What We Analyze
Each engine examines multiple independent signals in the text. The signals fall into four broad categories:
🔤 Vocabulary & Word Choice
Research shows that certain words increased dramatically in usage after large language models became widely available. We measure the density of these statistically anomalous vocabulary patterns — a technique grounded in peer-reviewed linguistic research.
📐 Sentence Structure & Rhythm
Human writers naturally vary their sentence lengths — short punchy sentences followed by longer complex ones. AI-generated text tends to have more uniform sentence length distribution. We measure statistical variation in sentence rhythm as a detection signal.
🔁 Repetition & Phrase Patterns
AI models generate text token-by-token and can unconsciously repeat similar phrases, sentence openings, or structural patterns throughout a piece. We analyze both exact and near-duplicate repetition across the text.
🧱 Structural & Stylistic Conventions
AI writing tends to follow predictable structural conventions — formal transitions, ritualistic conclusions, specific framing patterns. Human writing is less formulaic. We measure adherence to these AI-typical conventions.
How the Score is Calculated
Each signal produces an independent score. These are combined using a weighted model — signals with stronger research backing contribute more to the final result. The two engine scores are then merged based on which writing style best matches the input text.
The result is expressed as a percentage (0–100) representing the estimated probability that the text was AI-generated. We deliberately cap confidence below 100% — no browser-based detector can claim absolute certainty, and we believe in honest, transparent results.
100% Private — Runs in Your Browser
All detection happens client-side using JavaScript. Your text is never sent to any server. This means:
- Zero data storage — we cannot see what you paste
- Works offline after the page loads
- No account or login required — ever
- Results are instant — no API latency
Research Foundation
Our detection approach is informed by published academic research on AI-generated text, including studies on vocabulary drift in post-LLM corpora, statistical analysis of AI writing patterns, and human vs. AI text classification benchmarks.
Detection technology evolves alongside AI models. We continuously refine our algorithms as new models are released and writing patterns shift.