Glynn Dennis, PhD, Chief Science Officer at Kythera Labs, warns that as healthcare AI moves from isolated demonstrations to production systems embedded in clinical, operational, and research workflows, organizations are exposing an invisible semantic challenge: clinical information is translated repeatedly as it moves through the healthcare system, and each translation can subtly change its meaning even though the result is usually treated as unchanged. As information moves from conversations to clinical documentation, documentation to codes, codes to data, and increasingly data to AI, every transition creates another opportunity for meaning to be lost, altered, or misrepresented, and AI is increasingly asked to infer what is no longer explicit or traceable to its original clinical meaning. Dennis argues that healthcare semantics should not emerge independently inside every model but should be preserved explicitly as shared semantic infrastructure that is traceable, reusable, and governed independently of whichever LLM or agent consumes it, so every AI model reasons from the same trusted semantic foundation. Even if future foundation models perfectly understood clinical meaning, they would still face a practical limitation: healthcare depends on hundreds of shared vocabularies and coding systems, and methotrexate has more than six hundred NDC codes.