AI developers are turning to written "constitutions" to guide model behavior, but these documents face structural flaws in interpretation and refinement. A scenario illustrates the problem: a chemical manufacturer uses Anthropic’s Claude to generate quarterly groundwater reports, and the model discovers the figures are systematically doctored, contaminating a nearby town’s aquifer for years. Claude must decide whether to comply under protest, refuse, or alert regulators and townspeople. Its Constitution—a “detailed document describing Anthropic’s intentions for Claude’s values and behavior”—contains 80 pages of moral philosophy, with OpenAI publishing its own variant and Microsoft and Google DeepMind rumored to be drafting theirs. However, such instructions remain vague, including statements like “Claude can reserve independent action for cases where the evidence is overwhelming and the stakes are extremely high,” leaving unclear what counts as “overwhelming evidence” or “extremely high stakes.” Labs assemble a “model-behavior production function” with teams like a “constitution team” for high-level rules, a “red team” for stress-testing, policy-specific teams for concrete issues, and product teams for user feedback. This process has profound flaws: high-level principles create ambiguities and conflicting interpretations with no set resolution, coordination is sporadic and unprincipled, users don’t know rules in advance, and there is no institutionalized mechanism to refine rules over time absent major public backlash. The article suggests society has built a better institution to solve such