ai-detector

Open source application written in Python to check for AI content in paragraph input.

How to get started

Install dependencies

python -m venv venv && source venv/bin/activate
pip install transformers torch scikit-learn numpy textstat joblib datasets

Download the small backbone models

python ai_humaniser.py --download-models

Run interactively

python ai_humaniser.py

Paste a paragraph, hit ↵ twice, and you’ll get an AI-likelihood score plus a friendly verdict.

Benchmark against GPTZero

If you have a CSV with columns text,gptzero_score, load it in a notebook:

import pandas as pd
from ai_humaniser import Detector
det = Detector.load()
df = pd.read_csv("gptzero_eval.csv")
df["our_prob"] = df["text"].apply(det.predict_proba)
mape = det.mape_vs_gptzero(df["text"].tolist(), df["gptzero_score"].tolist())
print("MAPE vs GPTZero:", mape)

Tuning & extending

• Retrain with your own labelled data:

python ai_humaniser.py --retrain

The script automatically builds a new dataset (10 k human + 10 k AI paragraphs) and trains a logistic-regression classifier on the richer feature set in the code.

• Add extra features or swap in a different classifier – the Detector and _FeatureExtractor classes are cleanly separated; modify _FeatureExtractor.extract() or switch to an XGBoost model for potentially higher accuracy.
• API integration – wrap Detector.load() inside a Flask/FastAPI route to turn the script into a lightweight web service.