The Vatican's AI Policy Is Uncomfortably Good
Pope Leo XIV has issued what appears to be the first papal encyclical specifically addressing artificial intelligence. Magnifica Humanitas - 42,300 words, signed 15. May 2026., presented to the world on 25. May 2026. with the author himself in attendance - calls for AI to serve humanity rather than concentrate power in the hands of the few, and warns that opaque AI systems controlled by a small number of companies represent a risk of "new forms of dehumanisation."
It is worth pausing to note that a religious institution whose core texts were written in Aramaic, Greek, and Latin has produced governance thinking that is more structurally coherent than many documents produced by bodies specifically created for this purpose.
A secondary detail sharpens the picture. Christopher Olah - Anthropic's interpretability researcher, the person who has done more than almost anyone to make neural networks legible from the inside - was invited to speak at the encyclical's Vatican presentation, seated among cardinals and theologians. As Corey Quinn observed (via Simon Willison): Olah works for Anthropic. Anthropic runs its most intensive compute on SpaceX's COLOSSUS infrastructure, controlled by Elon Musk - a man whose intellectual mentor, Peter Thiel, wrote in 2009 that he no longer believes freedom and democracy are compatible. The safety researcher and the ideologically motivated infrastructure owner are separated by one corporate entity. Both, apparently, influenced the Pope. Through different doors.
The Right Question
The encyclical's central concern is not technical. It does not discuss benchmark scores, capability evaluations, or the relative merits of different regulatory frameworks. It asks a simpler and more fundamental question: in whose interest does AI operate, and who gets to decide?
This is not a new question. It is - however - arguably the central question in AI governance.
What the encyclical adds is a clarity of framing that policy documents often obscure under layers of technical specification: if AI systems are controlled by "a small number of entities," and if those entities are not democratically accountable, then... does the deployment of AI at scale represents a redistribution of power - from the many to the few - that no technical capability assessment can resolve?
The papal diagnosis does match what practitioners observe at ground level, and what a full accounting of the tech-right ideological programme makes explicit: the goal is not to build useful tools for humanity.
It is to achieve positions from which governance becomes irrelevant.
The encyclical's moral language and the governance analyst's structural language reach the same conclusion from different directions.
They do not cite each other.
They do not need to.
Compliance Is Not Accountability
The EU AI Act's risk-based framework is sophisticated and well-constructed. It creates obligations, conformity assessments, transparency requirements, and human oversight provisions. What it does not do - and it cannot do, within its legal structure - is address the structural question of who controls the physical and computational infrastructure on which AI runs.
A company can comply with every AI Act obligation while the data centres in which its models run are owned by actors with explicitly anti-democratic ideological commitments. A company can provide all required technical documentation while its training data practices are structured in ways that systematically exclude civil society from the audit process.
Compliance is what you demonstrate to the auditor.
Accountability is what you owe to the person whose pension fund, medical record, or deportation order the system processed.
A great deal of institutional energy goes into producing documentation that makes these look like the same thing. However, they are not - and the AI Act, as currently structured, addresses the first more reliably than the second. The financial mechanism by which this gap is exploited is already visible.
The Human Dignity Framing
The encyclical uses the language of "opaque AI" - systems that cannot be understood, questioned, or challenged by the people they affect. This is the AI governance problem stated from the perspective of human dignity rather than risk management. Risk management asks: "could this system hurt someone, and can we document that we checked?" Human dignity asks: "is this system arranged in ways that serve the people it affects?" The first question has a template. The second has not yet been reduced to one, though the drafting process is presumably ongoing. The Clause notes that human dignity has been assigned a project manager.
The effect is identical: systems that cannot be audited, that operate without meaningful oversight, that make decisions affecting lives without recourse or transparency - these are systems where power flows in one direction, permanently.
Simon Willison's observation from April - that "the people do not yearn for automation" - is the secular version of the same point. People want problems solved. They do not want agency replaced. When AI deployment is framed as automation imposed on users for the benefit of organisations, the human response is resistance. When it is framed as capability extended to individuals for their own benefit, the response is adoption. The distinction is not technical. It is moral - in the precise sense that the encyclical uses the word.
The papal framing adds a dimension that policy documents cannot easily incorporate: the question of whether a technology serves human flourishing is prior to the question of whether it is technically safe. A system can be technically safe - no data leaks, no bias above a given threshold, full audit logs - and still be arranged in ways that concentrate power, reduce human autonomy, and serve interests that are not those of the people it affects. A perfectly compliant system that serves the wrong master is still serving the wrong master.
The encyclical is impolite enough to say so.
Policy documents are not.
EU AI governance was built on fundamental rights rather than pure economic efficiency, which is its comparative advantage. The encyclical is a reminder of what that foundation requires in practice: that AI systems must be genuinely accountable to the people they affect, not merely compliant with the rules set by those who profit from them.
The document will not be cited in EU AI Board technical standards. Standards bodies cite prior standards, not encyclicals - this is a matter of institutional propriety rather than relevance. (The Clause notes, however - as obiter dictum, that good ideas tend to survive the drafting process by haunting the premises rather than appearing in the footnotes)
The question "in whose interest?" is one that technical governance frameworks tend to displace rather than answer. It does not go away because it is uncomfortable. It accumulates.