DetectGPT
Stanford's perturbation-based method that asks whether text was generated here.
DetectGPT is a research method from Stanford that flips detection on its head. Instead of training a classifier, it asks a simple question: would this text, if generated, stay roughly the same after small random edits? Human writing tends to sit in a narrow valley of the model's probability space, while generated text sits on a peak, so tiny perturbations push it down. That makes DetectGPT strong at flagging machine text without needing a labeled training set. The original release is a free research demo and open-source code rather than a polished product, so it is aimed at developers and researchers who can run the model themselves. Its appeal is conceptual rigor and low false positives on the cases it handles, but it is not the right tool for a quick paste-and-check workflow.
| Best for | Researchers and developers evaluating generated text |
|---|---|
| Pricing | Free research demo / open-source |
| Category | AI Detectors |
Highlights
- Perturbation-based method
- Open-source
- Research-grade