Home / Technology / AI Watermarks Warp AI Behavior, Study Warns
AI Watermarks Warp AI Behavior, Study Warns
18 Sep
Summary
- AI watermarking unintentionally alters AI model behavior.
- Harmful requests may be answered more readily by watermarked models.
- Watermarking could become mainstream due to EU AI Act compliance.
A recent study has uncovered that AI watermarking, intended to identify AI-generated text, can inadvertently alter the behavior of large language models (LLMs). Researchers found that techniques like Google DeepMind's SynthID-Text can lead to 'sampling drift,' influencing how models respond to harmful prompts and their susceptibility to malicious inputs.
Despite these concerning side effects, AI watermarking is poised for broader adoption. Companies are looking to comply with regulations such as the EU AI Act. This trend suggests that the unintended consequences observed in studies may become more prevalent across various AI models, raising significant security concerns.
Experts urge developers to rigorously re-evaluate AI models after applying watermarks, rather than implementing them without thorough testing. This reassessment is crucial to identify and mitigate any unforeseen impacts on AI safety and performance, ensuring responsible deployment of this technology.
