Temporal Discrepancy Tomography (TDT)

Official implementation of "AI-Generated Text is Non-Stationary: Detection via Temporal Tomography"

Paper | Demo

Abstract

The field of AI-generated text detection has evolved from supervised classification to zero-shot statistical analysis. However, current approaches share a fundamental limitation: they aggregate token-level measurements into scalar scores, discarding positional information about where anomalies occur. Our empirical analysis reveals that AI-generated text exhibits significant non-stationarity—statistical properties vary by 73.8% more between text segments compared to human writing. This discovery explains why existing detectors fail against localized adversarial perturbations that exploit this overlooked characteristic.

We introduce Temporal Discrepancy Tomography (TDT), a novel detection paradigm that preserves positional information by reformulating detection as a signal processing task. TDT treats token-level discrepancies as a time-series signal and applies Continuous Wavelet Transform to generate a two-dimensional time-scale representation, capturing both the location and linguistic scale of statistical anomalies.

Key Results

RAID Benchmark (Table 1 in paper)

• Overall: 0.855 AUROC (7.1% improvement over best baseline)
• Recipes: 0.875 AUROC (15.3% improvement)
• Poetry: 0.894 AUROC (8.1% improvement)
• News: 0.869 AUROC (13.3% improvement)

HART Benchmark (Tables 2-3 in paper)

• Level 1 (Simple Detection): 0.825 AUROC (5.8% improvement)
• Level 2 (Adversarial Paraphrasing): 0.812 AUROC (14.1% improvement)
• Level 3 (Humanization): 0.891 AUROC (2.4% improvement)

Cross-Model Generalization (Tables 4-5 in paper)

• QWEN-3-0.6B English: 0.724 AUROC (6.3% improvement)
• Spanish News: 0.638 AUROC (11.4% improvement at Level 1)
• Arabic News: 0.674 AUROC (33.1% improvement at Level 2)

Efficiency

• Only 13% computational overhead compared to scalar methods
• Maintains O(n log n) complexity

Method

TDT reformulates AI text detection as a signal processing task through three key stages:

1. Signal Generation

Token-level discrepancy scores Z(x) = [z₁, z₂, ..., zₙ] are converted to a continuous signal using Gaussian Kernel Density Estimation:

Z̃(x,t) = (1/nh) ∑ᵢ₌₁ⁿ K((t-i)/h) zᵢ

2. Continuous Wavelet Transform

The signal is decomposed using the Morlet wavelet to create a 2D time-scale representation:

W(a,b) = (1/√a) ∫ Z̃(x,t) ψ*((t-b)/a) dt

3. Multi-Scale Feature Extraction

Energy is extracted from three linguistically-motivated bands:

• Morphological (scales 1-4): Word-level anomalies
• Syntactic (scales 5-8): Phrase-level patterns
• Discourse (scales 9-12): Paragraph-level coherence

The final representation is:

S_TDT(x) = [‖W_morph‖_F, ‖W_syn‖_F, ‖W_disc‖_F]

Installation

# Python 3.12+ required
pip install -r requirements.txt

# For PDF report generation (optional)
# Ubuntu/Debian:
sudo apt-get install python3-cffi python3-brotli libpango-1.0-0 libpangoft2-1.0-0

# macOS:
brew install pango

# Note: WeasyPrint requires system dependencies for PDF generation

Quick Start

Interactive Demo with Professional PDF Reports

The demo script provides a comprehensive TDT analysis system with professional PDF report generation capabilities.

Usage Examples

# Run interactive mode (default)
python demo.py

# Analyze specific text and generate PDF report
python demo.py --text "Your text to analyze here" --output analysis_report.pdf

# Analyze text from file
python demo.py --file input.txt --output report.pdf

# Run example detections with automatic report generation
python demo.py --examples

# Run in interactive mode
python demo.py --interactive

# Analyze without generating PDF report
python demo.py --text "Your text" --no-report

# Custom detection threshold
python demo.py --text "Your text" --threshold 0.003

Command Line Options

• --text: Text to analyze directly
• --file: Path to file containing text to analyze
• --examples: Run example detections with professional reports
• --interactive: Run in interactive mode
• --output: Custom path for PDF report (default: tdt_report_.pdf)
• --no-report: Disable PDF report generation
• --threshold: Detection threshold (default: 0.0023)

Generated Report Features

The demo generates comprehensive PDF reports including:

• Executive Summary: Detection result, confidence score, and key metrics
• Text Analysis: Input text visualization with statistics
• Token-Level Visualization: Discrepancy signal with suspicious regions highlighted
• 2D Wavelet Scalogram: Time-scale representation showing linguistic anomalies
• Multi-Scale Features: Energy extracted from morphological, syntactic, and discourse bands
• Professional Layout: Two-page report with aligned visualizations

Reports are automatically saved to the specified output path or tdt_report_<timestamp>.pdf by default.

Run Experiments

# RAID benchmark
./test.sh

# HART benchmark
./main.sh

# Multilingual experiments
./langs.sh

Project Structure

TDT/
├── demo.py                    # Interactive TDT demonstration
├── requirements.txt           # Dependencies
├── scripts/                   
│   ├── delegate_detector.py   # Main detection pipeline
│   └── detectors/            
│       ├── t_detect.py       # TDT implementation
│       └── configs/          
│           └── tdt.json      # TDT configuration
├── benchmark/                 # Datasets (not included)
├── test.sh                   # RAID experiments
├── main.sh                   # HART experiments
└── langs.sh                  # Multilingual experiments

Models

TDT uses the following pre-trained models from HuggingFace:

• Reference Model: tiiuae/falcon-7b
• Scoring Model: tiiuae/falcon-7b-instruct

These will be automatically downloaded on first use.

Citation

@misc{west2025aigeneratedtextnonstationarydetection,
      title={AI-Generated Text is Non-Stationary: Detection via Temporal Tomography}, 
      author={Alva West and Yixuan Weng and Minjun Zhu and Luodan Zhang and Zhen Lin and Guangsheng Bao and Yue Zhang},
      year={2025},
      eprint={2508.01754},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.01754}, 
}

Key Contributions

1. Empirical Evidence: First to demonstrate that AI-generated text exhibits 73.8% higher non-stationarity than human writing
2. Novel Paradigm: Introduces temporal analysis to preserve positional information discarded by scalar methods
3. State-of-the-Art Performance: Achieves significant improvements across all benchmarks while maintaining efficiency

License

MIT License - see LICENSE file for details.