Fast-DetectGPT

February 7, 2026 · View on GitHub

This code is for ICLR 2024 paper "Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature", where we borrow or extend some code from DetectGPT.

Paper | LocalDemo | OnlineDemo | OpenReview

  • 1/31/2026: We find that Fast-DetectGPT with Llama3-8B/Llama3-8B-Instruct as the sampling/scoring models substantially outperforms falcon-7b/falcon-7b-instruct, especially on LRM-generated texts. We provide the demo online.
  • 9/13/2025: We have recover the demo at https://fastdetect.net.
  • :fire: API support is launched. Please check the API page in the demo.
  • :fire: Fast-DetectGPT can utilize GPT-3.5 and other proprietary models as its scoring model now via Glimpse.
  • :fire: Llama3-8B/Llama3-8B-Instruct > falcon-7b/falcon-7b-instruct >> gpt-j-6B/gpt-neo-2.7B, as the sampling/scoring models.

Brief Intro

Method 5-Model Generations ↑ ChatGPT/GPT-4 Generations ↑ Speedup ↑
DetectGPT 0.9554 0.7225 1x
Fast-DetectGPT 0.9887 (relative↑ 74.7%) 0.9338 (relative↑ 76.1%) 340x
The table shows detection accuracy (measured in AUROC) and computational speedup for machine-generated text detection. The white-box setting (directly using the source model) is used for detecting generations produced by five source models (5-model), whereas the black-box setting (utilizing surrogate models) targets ChatGPT and GPT-4 generations. AUROC results are averaged across various datasets and source models. Speedup assessments were conducted on a Tesla A100 GPU.

Environment

  • Python3.8
  • PyTorch1.10.0
  • Setup the environment: bash setup.sh

(Notes: our experiments are run on 1 GPU of Tesla A100 with 80G memory.)

Local Demo

Please run following command locally for an interactive demo:

python scripts/local_infer.py

where the default sampling and scoring models are both gpt-neo-2.7B.

We could use gpt-j-6B as the sampling model to obtain more accurate detections:

python scripts/local_infer.py  --sampling_model_name gpt-j-6B

An example (using gpt-j-6B as the sampling model) looks like

Please enter your text: (Press Enter twice to start processing)
Disguised as police, they broke through a fence on Monday evening and broke into the cargo of a Swiss-bound plane to take the valuable items. The audacious heist occurred at an airport in a small European country, leaving authorities baffled and airline officials in shock.

Fast-DetectGPT criterion is 1.9299, suggesting that the text has a probability of 82% to be machine-generated.

Workspace

Following folders are created for our experiments:

  • ./exp_main -> experiments for 5-model generations (main.sh).
  • ./exp_gpt3to4 -> experiments for GPT-3, ChatGPT, and GPT-4 generations (gpt3to4.sh).

(Notes: we share generations from GPT-3, ChatGPT, and GPT-4 in exp_gpt3to4/data for convenient reproduction.)

Citation

If you find this work useful, you can cite it with the following BibTex entry:

@inproceedings{bao2023fast,
  title={Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text via Conditional Probability Curvature},
  author={Bao, Guangsheng and Zhao, Yanbin and Teng, Zhiyang and Yang, Linyi and Zhang, Yue},
  booktitle={The Twelfth International Conference on Learning Representations},
  year={2023}
}