AI safety systems are increasingly deployed in high-stakes domains such as mental health, yet they are built on implicit assumptions about language that do not hold in practice. We argue that AI safety systems implicitly decide whose language counts as interpretable, and that risk expressed outside those norms is systematically overlooked. At the same time, AI systems can learn new language faster than their safety mechanisms can adapt to it. This creates a structural gap: while models may recognize emerging expressions, safety guardrails remain anchored to static linguistic norms. We introduce the concept of linguistic lag to describe the delay between the emergence of meaning within a community and its recognition by safety systems. Through examples and empirical observations, we show that semantically equivalent expressions of distress yield divergent safety outcomes when expressed in adolescent digital dialects. We argue that this is not a data coverage problem but a consequence of mismatched adaptation dynamics: language evolves through rapid, decentralized social diffusion, while safety systems evolve