Who Voted for This Model?

There is a complaint about AI that our industry keeps mishearing. When people say they do not trust these systems, we answer with accuracy: look, fewer errors than the human baseline, better calibration every quarter. The numbers improve and the distrust does not move, because the complaint was never about error rates. It is the older complaint, the one political philosophy has been processing for three centuries: decisions are being made about me, by a power I never agreed to, running rules I never saw. That is not an accuracy problem. It is a legitimacy problem, and legitimacy does not come from benchmarks.

Population-scale decisions, zero-person franchises

Every deployed model is now a small government. It decides who sees what, who gets flagged, whose claim sails through and whose gets the extra scrutiny. Every model is an opinion, and the opinion governs, which makes the honest question not whether the opinion is good but who was enfranchised in forming it. Walk the value chain and count the voters: a training corpus nobody consented to be in, objectives set by a product team, refusal policies drafted by a trust and safety group, all of it fine-tuned by whoever employs the labelers. Competent people, every one. Electorate size: zero.

Accuracy answers the wrong complaint

The standard industry response is that outcomes are what matter, and if the system is right more often than the human it replaced, rational people should prefer it. But we do not run societies that way, and never have. A brilliant ruler nobody chose is still the thing constitutions exist to prevent, not because brilliance is bad but because unaccountable brilliance has a bill that always arrives. People accept enormous imperfection from processes they consider legitimate, juries, elections, and reject near-perfection from processes they do not. The industry keeps trying to buy legitimacy with accuracy. The currency is not convertible.

What consent could even mean at model scale

The reflex objection: you cannot referendum a refusal policy; the public is not going to review training corpora. True, and not the point. Consent at scale never meant everyone deciding everything; it means the governed having real entry points into the rules. For models, the entry points are nameable. The values a system runs on, stated publicly in contestable language rather than buried in a system prompt. Mechanisms where affected people actually weigh the tradeoffs, deliberation, not sentiment scraping: a difference I would summarize as a franchise, not a survey. A path to contest a rule, used and answered in public. And an independent record proving the system actually runs the values it declares, because a constitution nobody can verify is a press release, which is where the values layer meets the evidence layer.

Why I build this rather than write papers about it

Legitimacy sounds like a philosopher's problem, but every piece of it decomposes into infrastructure: value elicitation that scales, deliberation mechanisms that resist capture, publication formats for machine-readable values, verification that closes the loop. That is Sanctity's work, the values layer, and I build it rather than theorize it for one reason: the systems are already governing. The election, so to speak, already happened, and nobody held it. Retrofitting legitimacy into deployed power is brutally hard, which is the strongest argument for building the franchise now, while the governments are still small.

The question that travels

You do not need my products to use the question. Next time a system makes a call about you, or your company ships one that makes calls about others, ask it plainly: who voted for this model? Trace the values to the people who set them, and count how many of the governed appear anywhere in the chain. The number is almost always zero, and saying the number out loud is how the number starts to change. Accuracy is a fine ambition for a tool. The moment the tool starts governing, the standard changes, and the industry that fails to notice will have the new standard written for it, by exactly the people it never consulted.

I hold the falsifiable version of this too: if a decade from now populations happily accept high-stakes model governance with no franchise, no contest rights, and no visible values, merely because accuracy stayed high, then legitimacy was not the missing property and I misread the century. I do not expect to be wrong. Every technology that quietly governed eventually faced its constitutional moment, and the ones that had built legitimacy in advance had far gentler moments than the ones that had built only accuracy.