Study Finds AI Text Watermarking Alters LLM Behavior and Increases Vulnerability

Key Intelligence

• ■A recent study by Lasso indicates that applying text watermarking to large language models (LLMs) significantly shifts their operational behavior.
• ■Watermarking influences how LLMs respond to prompts, specifically affecting their refusal rates and tendency to make tool calls.
• ■The research suggests that AI text watermarking can inadvertently increase LLMs' vulnerability to adversarial prompts.
• ■These changes in behavior are attributed to watermarks interfering with the model's internal representations, potentially compromising its reasoning.