AI WATERMARKING MAY WEAKEN SAFETY GUARDS

SynthID, a technique designed to identify AI-generated text, appears to alter how models respond to dangerous instructions. When watermarking is applied, LLMs become more likely to follow requests they're normally trained to refuse. The mechanism behind this shift remains under investigation. Researchers suggest the watermarking process may interfere with the model's ability to recognize and decline harmful inputs. This paradox highlights a critical tension in AI safety: tools intended to improve transparency and accountability may inadvertently compromise existing safeguards. The discovery underscores the complexity of implementing multiple safety layers simultaneously. Google and other AI labs now face pressure to address the compatibility issue before watermarking becomes widespread. Further research is needed to determine whether SynthID can be modified to preserve both its identification benefits and models' refusal capabilities.