AI watermarking could make LLM guardrail adherence unpredictable — and that could be a big problem for the EU AI Act

AI watermarking aims to prove the authenticity of any textNew study finds it also changes LLM behavior – and in a bad wayEU AI Act could mean more models have watermarks despite side effectsNew Lasso research has revealed that AI watermarking could actually unintentionally change how LLMs behave following the testing of Google DeepMind's... Weiterlesen