AI text detectors struggle with specific and formal human writing
Specificity Bias and the Wikipedia Paradox: An xAI Analysis of AI Text Detectors under Domain Shift
Computation and Language
Summary
Detecting AI-generated text is important, but tools designed for this often fail when faced with new types of writing. The authors tested AI text detectors on unusual news stories written by open-source AI models and found mistakes. These detectors sometimes rely on signals like calls to action common in AI texts, miss fake specific details inserted by AI, and wrongly flag very formal human writing as AI because it looks too uniform. This shows that current AI text detection methods can't fully trust statistics alone and need better understanding of meaning.
What this means in practice
- •For content moderation teams: Improve monitoring systems to better detect AI-generated content in specialized local news and conversational text types.
- •For digital marketing teams: Refine detection of AI-generated promotional content by recognizing call-to-action patterns in synthetic text.
Authors
Nico Morloher, Johannes Reschke, Mohamed Mady
Abstract
The widespread adoption of Large Language Models (LLMs) demands reliable detection of human-authored versus synthetic text. However, current Transformer-based detectors often experience severe performance degradation during domain shifts and operate as opaque classifiers. This paper investigates the robustness and interpretability of fine-tuned Transformer models (DistilBERT, RoBERTa) by evaluating them against a custom out-of-distribution (OOD) dataset of hyper-local news synthesized via open-source LLMs (Llama 3, Mistral, Gemma). To analyze model decision boundaries, we apply a tripartite Explainable AI (xAI) framework using SHAP. Our analysis reveals three key phenomena: First, detectors often rely on a "Call-to-Action" syndrome inherent to instruction-tuned LLMs rather than generic machine text. Second, we observe a "Specificity Bias," where LLMs successfully bypass detection by hallucinating specific named entities and mimicking conversational imperfection. Finally, we illustrate the "Wikipedia Paradox," wherein formal, objective human writing is frequently misclassified as AI due to its low statistical burstiness. These findings demonstrate that purely statistical detection paradigms face systematic limitations in highly structured human domains, emphasizing the critical need for semantically-aware, interpretable detection architectures.