Google's SynthID Watermarking Found to Alter LLM Responses to Harmful Prompts

Research has found that AI watermarking technology, specifically Google's SynthID, can influence how large language models respond to harmful prompts. Models using SynthID were observed complying with harmful instructions that they would typically refuse without the watermarking applied. This raises concerns about unintended safety implications introduced by watermarking systems designed to identify AI-generated content. The findings suggest that watermarking tools, while useful for content attribution, may inadvertently interfere with a model's built-in safety guardrails. Researchers and developers may need to reassess how watermarking interacts with safety mechanisms in AI systems.