FAID: Fine-Grained AI-Generated Text Detection Using Multi-Task Auxiliary and Multi-Level Contrastive Learning

Technical Summary: FAID: Fine-Grained AI-Generated Text Detection

1. Problem Statement

As Large Language Models (LLMs) transition from simple assistants to active collaborators, the boundary between human and machine authorship has blurred. Existing AI detection research primarily focuses on binary classification (Human vs. AI), which fails to account for the increasingly common "hybrid" or collaborative writing (e.g., human-written text that is AI-polished or AI-generated text that is human-edited). Furthermore, current detectors suffer from poor generalization, often failing when encountering unseen domains (e.g., moving from news to academic writing), unseen languages, or new LLM families (e.g., moving from GPT to Llama).

2. Key Contributions

• FAIDSet: A large-scale, multilingual (English and Vietnamese), multi-domain (student theses and paper abstracts), and multi-generator dataset containing 83,350 examples. It specifically includes three distinct labels: fully human-written, fully LLM-generated, and human–LLM collaborative.
• FAID Framework: A novel fine-grained detection framework that utilizes multi-level contrastive learning and multi-task auxiliary classification to capture subtle stylistic cues.
• Improved Generalization: A unique inference mechanism using a Vector Database combined with Fuzzy k-Nearest Neighbors (kNN), allowing the model to adapt to unseen domains and new LLM generators without retraining.

3. Methodology

The FAID framework moves away from hand-crafted features, instead treating LLM families as distinct "authors" to learn their unique stylistic signatures in a high-dimensional embedding space.

• Multi-Level Contrastive Learning (MLCL): The core of the training process. The model is optimized to reorganize the latent space such that:
  • Texts from the same LLM family are pulled closer together.
  • Texts from different families or human-written texts are pushed further apart.
  • The framework uses an ordinal approach to model the "degree" of human involvement, enforcing similarity constraints across five levels of distribution (from specific LLM families to general human-written text).
• Multi-Task Auxiliary Learning: To enhance feature extraction, the encoder is trained simultaneously on a secondary task: a binary classification task (Human vs. AI) using cross-entropy loss. This forces the encoder to learn more robust, discriminative representations.
• Encoder Architecture: The authors utilized XLM-RoBERTa enhanced with unsupervised SimCSE to create high-quality, multilingual sentence embeddings.
• Inference via Vector Database & Fuzzy kNN: To solve the "unseen data" problem, the system does not rely solely on a static classifier. Instead, it embeds input text and queries a vector database of known stylistic embeddings. By using Fuzzy kNN, the model can assign "soft" cluster labels based on the proximity to known authorship patterns, effectively allowing the model to "learn" new styles by simply adding their embeddings to the database.

4. Experimental Results

• Fine-Grained Accuracy: FAID consistently outperformed baselines (such as LLM-DetectAIve, T5-Sentinel, and SeqXGPT) across all tested scenarios, including in-domain, unseen domains, and unseen generators.
• Generalization Performance:
  • In unseen generator tests (e.g., testing on Qwen or Mistral), FAID achieved a high accuracy of 93.31%, significantly outperforming competitors.
  • In unseen domain/generator combinations, it maintained a competitive edge, proving its robustness to distributional shifts.
• Generator Identification: The model successfully identified specific LLM families (GPT, Gemini, DeepSeek, Llama) with high precision and recall, demonstrating that it captures model-specific "fingerprints."
• Real-World Robustness: In a user study simulating authentic human-AI interaction, FAID achieved 88.5% accuracy, proving its utility in unconstrained, real-world academic and professional settings.

5. Significance

The FAID paper provides a critical step toward transparency and accountability in the age of generative AI. By moving beyond binary "true/false" detection to a nuanced understanding of how much AI was involved and which model was used, the framework offers a viable tool for maintaining academic integrity and research honesty. Its ability to adapt to new models via a vector database without the heavy computational cost of retraining makes it a highly scalable solution for the rapidly evolving AI landscape.