In January 2026, Anthropic published a heavily revised version of Claude's Constitution: the privately authored framework governing how its AI reasons through open-ended situations. The effort is part of a broader search for moral grounding, one that has recently taken the company as far as the Vatican. The instinct to make Claude adhere to a set of general principles is understandable, even laudable. There is, however, a fundamental democratic problem. A constitution a company writes for its own product is one thing when that product is optional. It becomes something else when a constitutionally governed AI starts sitting inside welfare systems, advice desks, and government services people cannot meaningfully avoid. In a new paper, I call this line the Infrastructure Threshold, and I argue we're already drifting toward it through pilots, partnerships, and conveniences that each look sensible on their own. Once we cross the threshold, a private normative framework is quietly doing public work, and our usual governance tools can't adequately reach it. Legislation would be the most appropriate route. Procurement -- the point at which governments actually buy these systems -- offers a faster (if imperfect) one: attach democratic authorship conditions before the technology becomes too ordinary to remove. However we proceed, the question is one worth settling deliberately now, before institutional dependence settles it by default.