Care is irreducibly local. A caregiver in Munich documents that a resident is "pleasantly confused." A generic large language model — trained on internet data, optimized for Silicon Valley use cases, deployed from servers in Virginia — has no idea what this means. It might flag it as contradictory. But "pleasantly confused" is clinical shorthand, developed over decades by practitioners who needed to communicate cognitive status efficiently.
This is Semantic Fragility: the tendency of general-purpose AI to break when confronted with the specialized language of care. A caregiver in rural Georgia speaks differently than a caregiver in Geneva. A care note written in Tagalog carries different semantic weight than one written in German. Care is cultural. Care is linguistic.
The dominant AI systems are built in Silicon Valley, trained on internet-scale data, optimized for English, and deployed from centralized cloud infrastructure. They are remarkable engineering achievements. They are also structurally unsuited for care.
European care needs infrastructure built on European principles — sovereignty, subsidiarity, linguistic diversity, and data protection as architectural commitments, not compliance afterthoughts.