Fast-DetectGPT
Zero-shot detector for machine-generated text from an ICLR 2024 paper, with a local demo script and online API.
Research code that scores text by conditional probability curvature, with a local demo.
Fast-DetectGPT is the reference code for an ICLR 2024 paper on detecting machine-generated text without training a classifier. It suits researchers and technical users who want to run and inspect a detection method themselves.
What it is
It scores a passage by comparing the probability of its tokens against alternatives sampled by a language model, and reports a criterion and a probability that the text is machine-generated. The repo includes experiment scripts for five-model and GPT-3 to GPT-4 generations.
Who it's for
It suits people comfortable with Python and PyTorch who want a method they can read. It does not offer a dashboard, highlighting or document upload in the repo, though the authors host an online demo and API.
Key features
Conditional probability curvature
Scores text with a statistic the paper presents as 340x faster than DetectGPT.
Interactive local demo
Run scripts/local_infer.py, paste text and read a machine-generated probability.
Choice of scoring models
Defaults to gpt-neo-2.7B; the README lets you pass gpt-j-6B as the sampling model.
Shared generations
Includes GPT-3, ChatGPT and GPT-4 generations for reproducing experiments.
Online demo and API
The README links fastdetect.net, where an API page is documented.
Proprietary model scoring
The README says Glimpse lets it use GPT-3.5 and other proprietary models as scoring models.
Up and running in 3 steps
Our honest verdict
What's great
- Published at ICLR 2024, so the method is documented and citable.
- The code is short enough for a technical reader to inspect or extend.
- The authors report a Llama3-8B pairing that outperforms falcon-7b models on reasoning-model text.
- Free hosted demo and API exist if you do not want to run models.
Where it falls short
- The authors ran experiments on one 80 GB A100 GPU; larger models need serious hardware.
- The environment is Python 3.8 and PyTorch 1.10, which is old to set up today.
- There is no web interface, sentence highlighting or batch upload in the repository.
- Detection scores are probabilities, not proof, and accuracy figures come from the authors' own datasets.
Pricing & license
Fast-DetectGPT is free under the MIT licence. The README links a free online demo and API at fastdetect.net; the repo shows no paid plan, so the only cost is your own compute.
Open remote AI-training roles
767 open remote AI-training roles, updated November 14, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
Do I need a GPU for Fast-DetectGPT?
The authors ran experiments on a Tesla A100 with 80 GB. The README does not state a minimum for the local demo, whose default models are gpt-neo-2.7B.
Can Fast-DetectGPT detect ChatGPT text?
The README reports results for ChatGPT and GPT-4 generations using surrogate models and publishes those generations for reproduction.
Is there a hosted version?
The README links an online demo at fastdetect.net and an API page inside it.
