Adapting AI Text Detection to a Moving Target: Test-Time Adaptation Under Distribution Shift

A study published on arXiv examines a critical vulnerability of deployed AI text detection systems. Current approaches rely on training-time access to labeled datasets of human-written and AI-generated text, but face three types of distribution shift that occur continually post-deployment: adversarial humanization, new LLM releases with different generation patterns, and temporal drift in human writing styles.

Critically, labeled data for these new distributions is typically unavailable. The paper also notes that existing approaches fail to leverage a key signal — LLMs' own representations of text uncertainty — which could aid test-time adaptation.

Published on arXiv cs.CL on June 25, 2026, this work is relevant to the ongoing arms race between AI text generation and detection, highlighting the need for adaptive rather than static detection methods.