This research examines the vulnerability of LLM watermarking techniques to detection by end users through specially crafted prompts, challenging the assumption of imperceptibility.

• LLM watermarking helps detect AI-generated content while maintaining output quality
• Current watermarking methods can be identified by users through strategic prompt engineering
• This vulnerability creates a security trade-off for LLM providers between transparency and effectiveness
• The findings suggest providers need more robust, truly imperceptible watermarking solutions

For security professionals, this research highlights critical gaps in current AI content authentication methods and demonstrates how seemingly secure systems may be reverse-engineered by determined users.

Can Watermarked LLMs be Identified by Users via Crafted Prompts?