Part 1: What AI Detection Actually Is
AI text detection is the process of determining whether a piece of text was written by a human or generated by a large language model (LLM) like ChatGPT, Claude, or Gemini.
There are two fundamentally different approaches:
Measures linguistic patterns: vocabulary frequency, sentence length variation, transition density, closing rituals. Runs in a browser. Free. No server needed. Accuracy: 65–90% on unedited AI text.
Uses another language model to estimate perplexity and burstiness scores. More accurate (80–92%). Requires a server, API call, and usually a subscription. Examples: GPTZero, Turnitin AI.
This site uses pattern-based detection — because it can run privately in your browser at zero cost, and because the accuracy gap between the two approaches is smaller than vendors claim.
Part 2: The Science Behind Detection
The strongest research basis for AI text detection comes from three sources:
- Kobak et al. (Science Advances, 2025) — Identified vocabulary words that spiked 3–28× in academic papers after ChatGPT's launch. These aren't opinions — they're measured from millions of papers.
- GPTZero / Tian (2023) — Formalized perplexity and burstiness as detection metrics. Burstiness measures how much sentence length varies — humans are bursty, AI is flat.
- Wikipedia AI Cleanup Project (2024) — Real-world labeled dataset from Wikipedia editors identifying and removing AI content. Validated pattern thresholds at scale.
Part 3: The 12 Detection Signals
Our detector uses 12 signals grouped into two engines:
- Vocabulary density (Kobak words)
- Significance markers
- Sentence burstiness
- Negation framing
- Closing ritual
- Transition overuse
- Phrase loop detection
- Subject monotony
- Semantic circularity
- Duplicate sentences
- Temporal artifacts
- Structural monotony
Full explanation of each signal: How the Detector Works.
Part 4: What Detection Can and Cannot Do
Can do:
- Identify statistical patterns consistent with AI generation
- Give a probability estimate (not a verdict)
- Show you which specific patterns fired and why
- Screen a large volume of text quickly and privately
Cannot do:
- Prove definitively that a human or AI wrote a text
- Detect AI with certainty in texts under 100 words
- Reliably detect well-humanized AI text (50–65% accuracy)
- Replace human judgment in academic integrity decisions
Part 5: How to Use Detection Results Responsibly
A high AI probability score means: "this text has patterns statistically associated with AI generation." It does not mean: "this person cheated."
Formal academic writing, ESL writing, and highly structured business writing all trigger some AI signals in humans. Always use detection scores as a prompt for conversation, not as a verdict.
Ready to practice? Use the free detector and try it on your own writing to see what scores you get.