AI vs Human Code Detector

Machine learning project to classify source code as either AI-generated or human-written. It analyzes language patterns and stylistic signals across Python, Java, and JavaScript code.

The project includes:

• A training pipeline for multiple classifiers
• Saved model artifacts under model/
• A Streamlit web app for interactive prediction
• Dataset folders for Python, Java, and JavaScript samples

Features

• Binary classification: ai vs human
• Multiple models trained and evaluated:
  • Logistic Regression
  • Random Forest
  • Gradient Boosting
  • XGBoost
• Character-level TF-IDF feature extraction
• Ensemble voting in the app for final prediction
• Confidence display per model and overall result

Project Structure

MLProject/
  app.py
  train.py
  requirements.txt
  README.md
  data/
    python/
      ai/
      human/
    java/
      AI/
      Human/
    js/
      AI/
      Human/
  model/
    logistic.pkl
    randomforest.pkl
    gradientboost.pkl
    xgboost.pkl
    vectorizer.pkl
    labelencoder.pkl
    python/
    java/
    js/
  tableau_data/

Prerequisites

• Python 3.9+
• pip

Installation

1. Create and activate a virtual environment.

python -m venv .venv

Windows (Command Prompt):

.venv\Scripts\activate

2. Install dependencies.

pip install -r requirements.txt

Train Models

Run:

python train.py

This will:

• Load code samples from data/
• Build language-specific TF-IDF features
• Train models
• Print classification reports
• Save artifacts into model/ (and language subfolders)

Run the App

Start Streamlit:

streamlit run app.py

Then open the local URL shown in terminal (usually http://localhost:8501).

How to Use

1. Paste code into the text area.
2. Click Analyze Code.
3. Review:

• Overall prediction (AI-Generated or Human-Written)
• Ensemble confidence
• Per-model predictions and confidence values

Dataset Format

Each language folder should contain two classes:

• AI-generated code files
• Human-written code files

Expected extensions:

• Python: .py
• Java: .java
• JavaScript: .js

Notes

• Prediction quality depends on dataset quality, size, and balance.
• Results are probabilistic and should be treated as supportive signals, not absolute truth.
• If model files are missing, run python train.py before launching the app.

Common Issues

• ModuleNotFoundError: install dependencies with pip install -r requirements.txt.
• Model not found warning in app: ensure training completed and artifacts exist in model/.
• Streamlit command not found: use python -m streamlit run app.py.

Future Improvements

• Add more languages and larger datasets
• Improve feature engineering and model calibration
• Add explainability signals for why a prediction was made
• Export prediction logs and metrics dashboard