Python package for AI-generated text detection. Provides a high level api for easy adoption and more granular customization for advanced use cases.
Implement an AI Plagarism detector in 4 lines of Python:
from pydetectgpt import detect_ai_text
text = "text you want to check here"
result = detect_ai_text(text)
print("AI Generated" if result else "Human Written")On the first run it may some time to load the model from Hugging Face. After that it will be relatively fast.
You can also chose different Detection Methods, decision thresholds and use any transformers model for the logits:
from pydetectgpt import detect_ai_text
text = "text you want to check here"
result = detect_ai_text(text, method="fastdetectgpt", threshold=1.9, model="Qwen/Qwen2.5-1.5B")
print("AI Generated" if result else "Human Written")The default thresholds are:
"loglikelihood": -1.8,
"logrank": -0.8,
"detectllm": 2.14,
"fastdetectgpt": 1.9,
These were selected to minimize false positives (minimize saying its AI text when its not).
There is also a CLI wrapper:
pydetectgpt "Your text here""Detection Result: AI Generated" or "Detection Result: Human Written"
If you want just the 0/1 result (ex for scripting) use the -q flag:
pydetectgpt "Your text here" -q0 or 1
For a full list of args see cli.py
PyDetectGPT supports four detection methods, in order of effectiveness:
- FastDetectGPT (default): Implementation of Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text
- DetectLLM: Implementation of DetectLLM: Leveraging Log Rank Information for Zero-Shot Detection of Machine-Generated Text
- LogRank: Average log token rank
- LogLikelihood: Basic log likelihood of the text
- Fast-DetectGPT: Efficient Zero-Shot Detection of Machine-Generated Text (Bao et al., ICLR 2024)
- DetectLLM: Leveraging Log Rank Information for Zero-Shot Detection of Machine-Generated Text (Su et al., 2023)
MIT