This research introduces TH-Bench, the first comprehensive framework to evaluate methods that can make AI-generated text appear human-written to evade detection systems.

Key Findings:

• Creates a standardized evaluation system for measuring the effectiveness of evading attacks on machine-generated text detectors
• Develops a taxonomy of attack methods focused on humanizing AI-generated content
• Provides security researchers with tools to understand vulnerabilities in current detection systems
• Enables better testing of detector robustness against sophisticated evasion techniques

Security Implications: As AI text becomes increasingly prevalent, this research helps organizations strengthen defenses against deceptive uses of AI-generated content that might bypass detection systems, informing better security protocols for content validation.

TH-Bench: Evaluating Evading Attacks via Humanizing AI Text on Machine-Generated Text Detectors