Feature-Level Insights into Artificial Text Detection with Sparse Autoencoders

2025-03-05

AI Summary

Addressing the core challenges of poor generalization and weak interpretability in AI-generated text detection (ATD), this work pioneers the integration of sparse autoencoders (SAEs) into ATD. We apply SAEs to the residual stream of Gemma-2-2b to disentangle highly discriminative, cross-model-stable implicit features. We propose a multidimensional attribution framework that jointly leverages domain- and model-specific statistical analysis, feature-guided intervention, and LLM-assisted semantic interpretation—enabling identification of multiple highly sparse, strongly activated, and semantically interpretable features indicative of AI-generated text. Empirical analysis reveals systematic stylistic biases in modern LLMs under information-dense conditions. Our approach significantly improves ATD generalization to unseen texts and novel LLMs while enhancing decision transparency and interpretability.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models
• Computer Vision: Interpretability, Explainability, and Transparency

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Artificial Text Detection (ATD) is becoming increasingly important with the rise of advanced Large Language Models (LLMs). Despite numerous efforts, no single algorithm performs consistently well across different types of unseen text or guarantees effective generalization to new LLMs. Interpretability plays a crucial role in achieving this goal. In this study, we enhance ATD interpretability by using Sparse Autoencoders (SAE) to extract features from Gemma-2-2b residual stream. We identify both interpretable and efficient features, analyzing their semantics and relevance through domain- and model-specific statistics, a steering approach, and manual or LLM-based interpretation. Our methods offer valuable insights into how texts from various models differ from human-written content. We show that modern LLMs have a distinct writing style, especially in information-dense domains, even though they can produce human-like outputs with personalized prompts.

Problem

• Enhance interpretability in Artificial Text Detection using Sparse Autoencoders.
• Identify interpretable features from Gemma-2-2b residual stream for ATD.
• Analyze differences between LLM-generated and human-written texts.

Innovation

• Sparse Autoencoders extract interpretable features
• Analyze semantics using domain-specific statistics
• Identify distinct writing styles of LLMs