# DAMAGE: Detecting Adversarially Modified AI Generated Text

**Authors:** Elyas Masrour, Bradley Emi, Max Spero  
**arXiv:** [2501.03437](https://arxiv.org/abs/2501.03437)  
**Date:** 2025-01-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2501.03437v1/

## Key points

- Authors Elyas Masrour, Bradley Emi, and Max Spero focused on detecting adversarially modified AI-generated text
- Studied 19 different paraphrasing tools to evaluate their impact on original text's meaning and fidelity
- Identified a significant loophole in existing AI software's ability to detect humanized text effectively
- Developed a robust model capable of detecting humanized AI text with low false positive rate using a specific implementation
- Conducted an experiment targeting their own detector, demonstrating its strong resilience against targeted attacks
- Emphasized the importance of developing sophisticated models to combat adversarial modifications in AI-generated content

## Summary

In their paper titled "DAMAGE: Detecting Adversarially Modified AI Generated Text," authors Elyas Masrour, Bradley Emi, and Max Spero delve into the realm of , a novel category of online tools designed to rephrase and alter AI-generated text in a manner that enables them to bypass detection by AI software. The study conducted by the authors involved an examination of 19 different  and paraphrasing tools, with a focus on evaluating their impact on the original text's meaning and fidelity. The findings revealed a significant loophole in many existing , as they struggled to identify humanized text effectively. To address this issue, the authors developed a robust model capable of detecting humanized AI text while maintaining a low false positive rate through the implementation of a . Furthermore, they conducted an experiment where they targeted their own detector by training a finely-tuned model specifically optimized against its predictions. Remarkably, the authors demonstrated that their detector exhibited strong  capabilities, remaining resilient even when subjected to such targeted attacks. This highlights the importance of developing sophisticated  to combat adversarial modifications in AI-generated content effectively.