Who chooses the guardians of AI?

External institutions are needed but as the Romans would say: Who watches the watcher?

One of the most important questions in the AI debate is no longer whether these systems need rules. Of course they do. The deeper question is who has the authority to write those rules, and on what basis they can claim legitimacy.

That is what makes Nick Caputo’s latest paper so important. He takes seriously the idea that the “constitutions” governing large language models such as Claude are not just marketing language or internal policy documents, but real constitutions in an emerging sense: frameworks that shape what these systems are permitted to do, how they reason, and what kind of agents they become in practice. Once that point is granted, the question of legitimacy becomes unavoidable. If the constitution matters, then so does the authority behind it.

Caputo is right to insist that external institutions are needed. A company cannot simply write the rules for a powerful system, declare the result aligned, and expect the world to accept that as legitimate. There must be accountability, records that can be trusted, and structures outside the firm capable of evaluating whether dissent, endorsement, and governance mean anything at all. That is a serious and necessary insight.

But once we concede that point, we immediately run into the older and harder question the Romans posed long ago: Quis custodiet ipsos custodes? Who watches the watcher?

That question has become especially urgent in a world crowded with agenda-driven NGOs, politicized “experts,” and institutions that often claim neutrality while quietly advancing very specific anthropologies and moral visions.

If “external institutions” are to confer legitimacy on AI, who chooses them? Who authorizes them? Who judges whether they serve the common good rather than the ideology of the class that staffs them? The problem of governance does not disappear when you move outside the corporation. It merely changes location.

This is where Catholic realism is indispensable. As a Catholic, I welcome the growing recognition that AI cannot be left to private companies alone. We should be grateful whenever serious voices insist, as the Holy See continues to insist, on robust governance. That phrase matters. Governance is necessary because these systems will shape human life at enormous scale, and because power without accountability always becomes dangerous.

But Catholic realism also refuses to romanticize the governors. Every institution, public or private, is staffed by fallen human beings. Every structure of oversight is tempted by ideology, capture, vanity, and the will to dominate.

So the challenge is not merely to demand “external institutions,” as if the adjective solved the problem. The real challenge is to build institutions worthy of trust: institutions disciplined by transparency, subsidiarity, moral seriousness, and a truthful account of the human person. That last point matters most. No governance regime for AI will be legitimate if it is built on a false anthropology. If it cannot say clearly what a human being is, what dignity means, what responsibility requires, and why no machine may displace the moral primacy of the person, then its authority will always be shaky.

Caputo is right to press the legitimacy problem. But the Roman question remains. If AI is to be watched, then the watchers will have to be watched as well. And that means the final issue is not only governance, but the moral quality of the civilization doing the governing. In the end, the question is not just who writes the constitution for the machine. It is who forms the conscience of the people who claim the right to do so.