GPTZero
GPTZero is an AI text detection tool built by Edward Tian in 2023, widely used in academic settings to detect AI-generated submissions. It measures perplexity and burstiness to estimate AI probability.
Definition¶
GPTZero is an AI text detection tool developed by Edward Tian, then a Princeton University student, and released in January 2023. It became one of the most widely adopted detection tools in academic and educational settings following widespread concern about AI-generated academic submissions.
Detection Method¶
GPTZero's core detection signals are:
- Perplexity — measures how predictable each word choice is. AI-generated text has lower perplexity (more predictable word choices) than typical human writing.
- Burstiness — measures variation in sentence length. Human writing is more bursty (varied); AI text is more uniform.
GPTZero applies these signals through a trained classifier that produces both a document-level AI probability score and sentence-level highlighting showing which parts triggered detection.
Use Cases¶
- Educators checking student submissions for AI-generated content
- Researchers studying AI text generation
- Writers verifying their own humanized content before submission
Limitations¶
- False positives: Dense academic writing, ESL text, and consistent formal styles are sometimes flagged as AI-generated. Published studies suggest false positive rates of 4–20% depending on writing type.
- Evasion: Humanized text that has higher perplexity and burstiness falls below GPTZero's detection thresholds.
- Model dependency: Accuracy varies by the model used to generate the text; GPTZero performs best on GPT-3.5 and GPT-4 output.
- Not definitive: GPTZero's own documentation recommends treating scores as probabilistic signals, not proof of misconduct.
Pricing¶
GPTZero offers a free tier with limited scans and a paid plan for higher-volume use. API access is available for institutional integrations.