DAMASHA: Detecting AI in Mixed Adversarial Texts via Segmentation with Human-interpretable Attribution

TL;DR

This paper presents Info-Mask, a novel framework for detecting and segmenting mixed human-AI authored texts, incorporating stylometric cues, adversarial robustness, and human-interpretable attributions to improve trust and oversight.

Contribution

We introduce Info-Mask, a new method combining stylometric, perplexity, and boundary modeling for accurate mixed-authorship segmentation, along with an adversarial benchmark dataset MAS.

Findings

Info-Mask improves segmentation robustness against adversarial attacks.

The framework provides human-interpretable attributions for boundary decisions.

Our system establishes new performance baselines in mixed-authorship detection.

Abstract

In the age of advanced large language models (LLMs), the boundaries between human and AI-generated text are becoming increasingly blurred. We address the challenge of segmenting mixed-authorship text, that is identifying transition points in text where authorship shifts from human to AI or vice-versa, a problem with critical implications for authenticity, trust, and human oversight. We introduce a novel framework, called Info-Mask for mixed authorship detection that integrates stylometric cues, perplexity-driven signals, and structured boundary modeling to accurately segment collaborative human-AI content. To evaluate the robustness of our system against adversarial perturbations, we construct and release an adversarial benchmark dataset Mixed-text Adversarial setting for Segmentation (MAS), designed to probe the limits of existing detectors. Beyond segmentation accuracy, we introduce Human-Interpretable Attribution (HIA overlays that highlight how stylometric features inform boundary predictions, and we conduct a small-scale human study assessing their usefulness. Across multiple architectures, Info-Mask significantly improves span-level robustness under adversarial conditions, establishing new baselines while revealing remaining challenges. Our findings highlight both the promise and limitations of adversarially robust, interpretable mixed-authorship detection, with implications for trust and oversight in human-AI co-authorship.