The Science of AI Detection
How do we know if a text was written by a human or a machine? We look for the patterns that ChatGPT can’t hide.
The 3-Step Detection Process
Scan & Tokenize
We break your essay down into thousands of small units called “tokens” (words and sub-words) to analyze structure.
Pattern Matching
We compare the text against the known statistical patterns of LLMs like GPT, Gemini, and Claude.
Probability Score
We calculate the likelihood that the text was generated by AI and provide a final percentage score.
Two Key Metrics We Analyze
1. Perplexity
Definition: A measure of how unpredictable a text is.
AI models are trained to be “safe” and predictable. They choose the most statistically probable next word. Humans, however, are chaotic. We use unexpected words, slang, and creative metaphors.
- 📉 Low Perplexity: Likely AI (Predictable)
- 📈 High Perplexity: Likely Human (Creative)
2. Burstiness
Definition: The variation in sentence structure and length.
Humans tend to write in “bursts”. We might write one long, complex sentence followed by a short one. Like this. AI models tend to produce sentences of very uniform length and structure, creating a monotonous rhythm.
- 🤖 AI: Monotone rhythm (Robot-like)
- 👨🎓 Human: Dynamic rhythm (Bursty)
See the Difference
“Artificial intelligence is a rapidly growing field. It has many applications in various industries. Companies are using it to improve efficiency. It can handle data analysis very quickly. However, there are some concerns about privacy.”
Analysis: Sentences are all similar length. Simple Subject-Verb-Object structure. Very predictable.
“Everyone’s talking about AI these days, right? It’s exploding everywhere—from healthcare to finance. While companies love the efficiency boost (who doesn’t like saving time?), there’s a elephant in the room: privacy.”
Analysis: Varied sentence length. Use of rhetorical questions and idioms (“elephant in the room”).