How AI Detection Works
Understanding the multi-signal statistical ensemble that powers Cluster AI's GSM AI Detector.
1. Multi-Signal Feature Analysis
Our detector analyzes text across three independent analytical branches:
- Stylometric Heuristics: Evaluates sentence length variance (burstiness), vocabulary tier density, passive voice frequency, and grammatical uniformity.
- Multi-LLM Lexical Fingerprints: Identifies transitional phrases and token patterns characteristic of ChatGPT, Claude, Grok, Gemini, and Llama.
- Supervised ML Ensemble: Ensembled TF-IDF character and word n-gram classifiers with calibrated Logistic Regression and Random Forest models.
2. Human-First Prior Calibration
To prevent false accusations on formal academic papers, the model applies a human-first baseline prior ($P_0 = 0.05$) and dampens the machine learning probability when human stylistic markers (high sentence variety, informal idioms, first-person narrative) are present.