AI-FAKER

A Comprehensive Benchmark for Detecting, Attributing, and Explaining AI-Generated Images and Text

AI-Faker is a large-scale multimodal dataset introduced in the paper
Could AI Trace and Explain the Origins of AI-Generated Images and Text?
It comprises over 280,000 samples spanning different generative AI models:

• AI-Generated Images vs. AI-Modified Images
• AI-Text Completion vs. AI-Generated Peer Reviews
• General Use Cases vs. Malicious Use Cases

AI-Faker is intended to be the go-to dataset for model attribution, explanation, and robust forgery detection across multiple modalities.

---

Table of Contents

• Key Features
• Installation
• Dataset Overview
• Comparison
• Loading the Dataset
• Citation
• License

---

Key Features

• Multimodal Coverage: Spans both visual (fully and partially AI-generated images) and textual (short text completions and long-form peer reviews) content.
• General vs. Malicious Use Cases: Investigates standard generative tasks (e.g., text-to-image) alongside high-stakes misuse (e.g., face-swapping, AI-written academic reviews).
• Model Diversity: Includes outputs from popular LMMs (Midjourney, DALL-E, Stable Diffusion, etc.) and LLMs (GPT-4, Claude, LLaMA, etc.).
• Explainability Benchmark: Offers a unique testbed for explanation generation—whether an AI model (e.g., GPT-4) can explain why an image or text is attributed to a specific generator.

---

Installation

You can clone this repository and install dependencies from requirements.txt:

git clone https://github.com/your-username/AI-Faker.git
cd AI-Faker
pip install -r requirements.txt

Dataset Overview

Subset Description # Instances Category

• AI-Generated Images Fully diffused from text prompts (DALL-E, Midjourney). 50,000+ General

• AI-Modified Images Partially modified (face-swapped) 28,551 Malicious

• AI-Text Completion Short text completions (GPT-4, Cohere, LLaMA...) 50,000+ General

• AI-Paper Review Full-length peer reviews generated by LLMs 70,000 Malicious

Comparison:

Download the dataset

	Size	Domains	Dataset Size	GenAI Models
AI_generated_images	36.43Gb	Visual (natural + stylized prompts)	60,000 images (10K real, 50K AI-generated)	Midjourney, DALL-E, SDXL, Stable Diffusion, FLUX
AI_modified_images	1.69Gb	Multimedia (faces from movies, TikTok, YouTube)	28,551 images (6K real + 22.5K swapped)	Inswapper, SimSwap, UniFace, BlendSwap
AI_text_completion	36.3Mb	News, Books, Abstracts, Reviews, Reddit, Recipes, Wikipedia, Poetry	60,000 text samples (10K per LLM + human)	GPT-4, Cohere, LLaMA, Mistral, MPT
AI_paper_review	127.6Mb	Academic Paper Reviews (OpenReview)	70,000 text samples (10K per LLM + human)	GPT-4o, Claude 3.5, Gemini 1.5, DeepSeek, LLaMA-3, Mistral

Citation

If you find AI-Faker or its accompanying code valuable for your research, please cite our paper:

@misc{fang2025aitraceexplainorigins,
      title={Could AI Trace and Explain the Origins of AI-Generated Images and Text?}, 
      author={Hongchao Fang and Yixin Liu and Jiangshu Du and Can Qin and Ran Xu and Feng Liu and Lichao Sun and Dongwon Lee and Lifu Huang and Wenpeng Yin},
      year={2025},
      eprint={2504.04279},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2504.04279}, 
}