Human vs AI Text Classifier

Portfolio project

Human vs AI Text Classifier

This project builds a binary text classification system to distinguish between human-written and AI-generated text using a custom-labeled dataset. By combining TF-IDF vectorization with multiple machine learning models, it captures subtle linguistic patterns and style differences across writing sources.

Key Highlights

Custom Dataset: 5,000 samples (2,500 human + 2,500 AI-generated), curated and balanced by the author.

Models Trained:

Logistic Regression (Top performer: 100% accuracy)

Random Forest

Multinomial Naive Bayes

Calibrated Linear SVC

Text Preprocessing:

Lowercasing

Punctuation removal

Token cleaning

Evaluation Metrics:

Confusion Matrices

ROC Curves & AUC Scores

Precision, Recall, F1-Score

Visual Interpretations:

Word Clouds

Feature Importance Bar Charts

Model Accuracy Comparison

Deployment Ready:

Final model serialized with joblib

Predicts new input instantly

Results Summary

Model Accuracy AUC Score Logistic Regression 1.000 1.00 Linear SVC 0.998 - Naive Bayes 0.997 1.00 Random Forest 0.992 1.00

TF-IDF feature weights revealed strong interpretability.

Words like “industry”, “AI”, “intelligence” were predictive of AI text.

Words like “my”, “was”, “weather” were highly human-like.

ROC curves confirmed excellent model separability.

Note: Perfect accuracy is expected due to the high separability and structured nature of the dataset (intended for educational use).

Project Structure

HUMAN_VS_AI_CUSTOM/
│
├── data/
│   └── your_dataset_5000.csv
│
├── model/
│   ├── model_info.txt
│   └── text_classifier_5000.joblib
│
├── notebooks/
│   ├── Human vs AI Custom Dataset.ipynb
│   └── human_vs_ai_text_classifier.ipynb
│
├── requirements.txt
└── README.md