Human Texts Are Outliers: Detecting LLM-generated Texts via Out-of-distribution Detection

2025-10-07

AI Summary

This work addresses the poor generalization of AI-generated text detection by reframing human-written text as out-of-distribution (OOD) samples—a paradigm shift from conventional binary classification. Methodologically, we propose an end-to-end OOD detection framework leveraging single-class learning models (e.g., DeepSVDD, HRN) combined with energy-based scoring functions. Evaluated on the DeepFake dataset, our approach achieves 98.3% AUROC and AUPR, with an FPR@95 of only 8.9%. Crucially, it demonstrates strong cross-lingual, adversarial, and cross-model robustness—generalizing effectively to unseen large language models (e.g., GPT-4, Claude, DeepSeek) and multilingual settings. Our core contribution is the first systematic reformulation of AI-text detection as an OOD problem, which significantly enhances generalization to unknown models and domains while preserving detection accuracy.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Generation
• Computer Vision: Diffusion Models for Vision

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

The rapid advancement of large language models (LLMs) such as ChatGPT, DeepSeek, and Claude has significantly increased the presence of AI-generated text in digital communication. This trend has heightened the need for reliable detection methods to distinguish between human-authored and machine-generated content. Existing approaches both zero-shot methods and supervised classifiers largely conceptualize this task as a binary classification problem, often leading to poor generalization across domains and models. In this paper, we argue that such a binary formulation fundamentally mischaracterizes the detection task by assuming a coherent representation of human-written texts. In reality, human texts do not constitute a unified distribution, and their diversity cannot be effectively captured through limited sampling. This causes previous classifiers to memorize observed OOD characteristics rather than learn the essence of `non-ID' behavior, limiting generalization to unseen human-authored inputs. Based on this observation, we propose reframing the detection task as an out-of-distribution (OOD) detection problem, treating human-written texts as distributional outliers while machine-generated texts are in-distribution (ID) samples. To this end, we develop a detection framework using one-class learning method including DeepSVDD and HRN, and score-based learning techniques such as energy-based method, enabling robust and generalizable performance. Extensive experiments across multiple datasets validate the effectiveness of our OOD-based approach. Specifically, the OOD-based method achieves 98.3% AUROC and AUPR with only 8.9% FPR95 on DeepFake dataset. Moreover, we test our detection framework on multilingual, attacked, and unseen-model and -domain text settings, demonstrating the robustness and generalizability of our framework. Code, pretrained weights, and demo will be released.

Problem

• Detecting AI-generated texts via out-of-distribution detection
• Treating human texts as outliers versus machine-generated content
• Improving generalization across domains and unseen models

Innovation

• Reframing detection as OOD problem
• Using one-class learning methods like DeepSVDD
• Applying energy-based score learning techniques