🔍 AI vs Human Text Detector

A Production-Ready Machine Learning System to Detect AI-Generated Content with 87% Accuracy

---

📖 Table of Contents

1. Problem Statement
2. Why This Matters
3. Dataset Overview
4. Model Selection & Comparison
5. Why Logistic Regression?
6. Technical Architecture
7. Feature Engineering
8. Model Performance
9. Installation & Usage
10. Live Demo
11. Future Improvements

---

🎯 Problem Statement

Can we reliably distinguish between text written by humans and text generated by AI models like GPT-4, Claude, or Gemini?

With the rapid advancement of Large Language Models (LLMs), AI-generated content has become increasingly sophisticated and harder to detect. This poses challenges for:

• Academic Integrity - Detecting AI-written essays and assignments
• Content Moderation - Identifying AI-generated spam and fake news
• Trust & Authenticity - Verifying human-generated content online

This project builds a traditional Machine Learning classifier (not Deep Learning) to solve this binary classification problem using TF-IDF features and a Logistic Regression model.

---

⚠️ Why This Matters

Domain	Challenge	Our Solution
Education	Students using AI to write essays	Detect AI-generated submissions
Journalism	AI-written news articles	Verify authentic reporting
SEO Content	AI-generated blog spam	Filter low-quality content
Social Media	Bot-generated comments	Identify automated accounts

---

📊 Dataset Overview

Dataset 1: AI vs Human Content (20,000 rows, 13 columns)

Column	Description	Type
content	The actual text to analyze	Text
label	Target variable ('human' or 'ai')	Category
source	Content source (news, blog, academic, etc.)	Category
topic	Subject matter	Category
word_count	Number of words	Numeric
char_count	Number of characters	Numeric
complexity_score	Readability score (1-10)	Numeric
ai_model	Which AI generated it (NaN for human)	Category

Dataset 2: Sentence Dataset (9.8 million rows)

• Extracted sentences from essays
• Used for final model training
• Provides 500K sampled sentences for balanced training

Key Insight from EDA

• Correlation between numeric features and label:
• word_count: 0.05 (very weak)
• char_count: 0.07 (very weak)
• complexity_score: 0.001 (almost zero)
• Conclusion: Numeric features DON'T help prediction.
• Only TEXT content matters!

---

🧠 Model Selection & Comparison

I tested 4 different machine learning models on 437,000 training samples to find the best performer:

Model	Accuracy	Training Time	F1-Score	Precision	Recall
Logistic Regression	87.12%	8.5 sec	0.87	0.87	0.87
RidgeClassifier	87.12%	3.5 sec	0.87	0.87	0.87
MultinomialNB	84.75%	0.2 sec	0.85	0.84	0.85
SGDClassifier	83.96%	1.4 sec	0.84	0.84	0.84

Model Details

Model	Type	Pros	Cons
Logistic Regression	Linear	Fast, interpretable, works well with sparse data	May underfit complex patterns
RidgeClassifier	Linear	Similar to LR, slightly faster	No probability outputs
MultinomialNB	Naive Bayes	Very fast, good for text	Assumes feature independence
SGDClassifier	Linear	Handles large datasets well	Requires careful tuning

---

✅ Why Logistic Regression?

After extensive testing, Logistic Regression was chosen as the final model for several reasons:

1. Performance

Accuracy: 90.50% Precision: 87% (when model says AI, it's right 87% of the time) Recall: 74% for AI, 94% for Human F1-Score: 0.90

2. Interpretability

Unlike black-box models (Neural Networks, Random Forests), Logistic Regression provides:

• Feature importance (which words matter most)
• Clear decision boundary
• Easy to debug and explain

3. Speed

• Training time: 8.5 seconds on 437,000 samples
• Inference time: < 1 second per text
• Suitable for real-time applications

4. Sparse Data Handling

TF-IDF creates sparse matrices (95% zeros). Logistic Regression handles sparse data efficiently.

5. Probability Outputs

Provides confidence scores (0-100%) for each prediction.

6. No Overfitting

With L2 regularization, Logistic Regression generalizes well to unseen data.

7. Production Ready

• Small model size (~5MB)
• Easy to deploy
• Works without GPU

---

🔧 Feature Engineering

TF-IDF Vectorization Parameters

Parameter	Value	Why
max_features	10,000	Keep only top 10,000 most important words
ngram_range	(1,2)	Capture single words AND word pairs (e.g., "not good")
stop_words	'english'	Remove common words like 'the', 'and', 'is'
min_df	2	Ignore words that appear less than twice
max_df	0.95	Ignore words that appear in >95% of texts

Why TF-IDF?

Method	What it does	Why used
Term Frequency (TF)	How often a word appears	Captures word importance in document
Inverse Document Frequency (IDF)	How rare a word is across all documents	Reduces weight of common words
TF-IDF Score	TF × IDF	High score = important word for this document

Example TF-IDF in Action

• Document 1: "The cat sat on the mat"

• Document 2: "The dog sat on the log"

• Word 'cat' appears only in Doc1 → High IDF → High importance

• Word 'the' appears in both docs → Low IDF → Low importance

🚀 Installation & Usage

Prerequisites

Python 3.13+
pip package manager

# 1. Clone the repository
git clone https://github.com/Nerdy37/AI-Text-Detector.git
cd AI-Text-Detector

# 2. Install dependencies
pip install -r requirements.txt

# 3. Run the application
python app.py

# 4. Open browser to http://localhost:7860

live demo

![veritext]