Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Explaining Generalization of AI-Generated Text Detectors Through Linguistic Analysis

Yuxi Xia, Kinga Stańczak, Benjamin Roth

TL;DR

AI-text detectors exhibit near-perfect in-domain accuracy but fail to generalize across prompts, models, and domains. The authors propose a linguistic-analysis framework and build a large, controlled benchmark with 6 prompting strategies, 7 LLMs, and 4 domains, then fine-tune RoBERTa- and DeBERTa-based detectors to study cross-domain generalization. They quantify shifts across 80 linguistic features and find significant correlations, with features like pronoun usage, verb tense, and passive voice predicting generalization gaps in some settings. The results offer interpretable signals to diagnose and improve detector robustness, highlighting that evaluation must extend beyond in-domain performance for real-world deployment.

Abstract

AI-text detectors achieve high accuracy on in-domain benchmarks, but often struggle to generalize across different generation conditions such as unseen prompts, model families, or domains. While prior work has reported these generalization gaps, there are limited insights about the underlying causes. In this work, we present a systematic study aimed at explaining generalization behavior through linguistic analysis. We construct a comprehensive benchmark that spans 6 prompting strategies, 7 large language models (LLMs), and 4 domain datasets, resulting in a diverse set of human- and AI-generated texts. Using this dataset, we fine-tune classification-based detectors on various generation settings and evaluate their cross-prompt, cross-model, and cross-dataset generalization. To explain the performance variance, we compute correlations between generalization accuracies and feature shifts of 80 linguistic features between training and test conditions. Our analysis reveals that generalization performance for specific detectors and evaluation conditions is significantly associated with linguistic features such as tense usage and pronoun frequency.

Explaining Generalization of AI-Generated Text Detectors Through Linguistic Analysis

TL;DR

AI-text detectors exhibit near-perfect in-domain accuracy but fail to generalize across prompts, models, and domains. The authors propose a linguistic-analysis framework and build a large, controlled benchmark with 6 prompting strategies, 7 LLMs, and 4 domains, then fine-tune RoBERTa- and DeBERTa-based detectors to study cross-domain generalization. They quantify shifts across 80 linguistic features and find significant correlations, with features like pronoun usage, verb tense, and passive voice predicting generalization gaps in some settings. The results offer interpretable signals to diagnose and improve detector robustness, highlighting that evaluation must extend beyond in-domain performance for real-world deployment.

Abstract

AI-text detectors achieve high accuracy on in-domain benchmarks, but often struggle to generalize across different generation conditions such as unseen prompts, model families, or domains. While prior work has reported these generalization gaps, there are limited insights about the underlying causes. In this work, we present a systematic study aimed at explaining generalization behavior through linguistic analysis. We construct a comprehensive benchmark that spans 6 prompting strategies, 7 large language models (LLMs), and 4 domain datasets, resulting in a diverse set of human- and AI-generated texts. Using this dataset, we fine-tune classification-based detectors on various generation settings and evaluate their cross-prompt, cross-model, and cross-dataset generalization. To explain the performance variance, we compute correlations between generalization accuracies and feature shifts of 80 linguistic features between training and test conditions. Our analysis reveals that generalization performance for specific detectors and evaluation conditions is significantly associated with linguistic features such as tense usage and pronoun frequency.
Paper Structure (41 sections, 5 equations, 4 figures, 7 tables)

This paper contains 41 sections, 5 equations, 4 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Datasets for AI Text Detection.
  4. AI Text Detection Models.
  5. Generalization in Detection.
  6. Explaining Generalization Behavior.
  7. Dataset Creation
  8. Human-written text
  9. Data cleaning for human-written text.
  10. AI Text Generated with LLMs
  11. LLMs.
  12. Prompts.
  13. Data Cleaning for AI-Generated Text.
  14. Generalization of AI-Text Detectors
  15. Generalization Testing
  16. ...and 26 more sections

Figures (4)

  • Figure 1: Comparison of aggregated generalization performance and aggregated feature shifts across all evaluation settings. Similar patterns in the two heatmaps indicate that certain feature shifts are correlated with reduced generalization accuracy.
  • Figure 2: Cross-prompt generalization and feature shifts when evaluating on a specific model and dataset. A clearer correlation is observed than the overall cross-prompt correlation in Figure \ref{['fig:main-corr-heatmap']}.
  • Figure 3: The workflow to get the overall cross-prompt generalization and feature shift results.
  • Figure 4: The more detailed comparison of different linguistic features across different configurations as well as the differences between human and AI text. We present the features that have the strongest correlations for different dimensions (underlined features in Table \ref{['fig:main-corr']} in the main paper).