On May 25, the Vatican published its first encyclical on artificial intelligence. On stage for the presentation stood Chris Olah, co-founder of Anthropic and head of the company's interpretability research. He asked for critics.
No Western head of government commented. No secular regulatory authority referenced the document. No tech op-ed page read it cover to cover.
It is a mistake.
Magnifica Humanitas spans 245 paragraphs, signed on May 15, marking the 135th anniversary of Rerum Novarum. It addresses AI, autonomous weapons, children's mobile usage, the labor market, and disinformation. It is easily misread as a confessional text without operational relevance for a secular society. And it is read wrong.
Because it is not a pastoral letter. It is a policy document.
It is not religion. It is classification.
Papal encyclicals have a history. Rerum Novarum (1891) took labor from being a mere commodity on a market and moved it into the sphere of human dignity. Centesimus Annus (1991) did the same with the market economy: yes, but subordinated. Laudato Si’ (2015) did it with the biosphere.
This is not about theology in any broad sense. It is about what counts as unconditional private property, and what counts as something else.
Magnifica Humanitas executes the exact same maneuver with AI infrastructure. It is the document’s only important move. Everything else is a consequence.
Three paragraphs do the heavy lifting
Out of the 245 paragraphs, three carry the structural weight.
Paragraph 9: Technology is never neutral. It bears the imprint of those who design, finance, regulate, and use it. It sounds banal. It is not. It is a direct contradiction of the techno-libertarians' standard position that the tool can be separated from the user. That position underpins Section 230, the defense of platform neutrality, and the "we're just the rails" rhetoric. If Leo's paragraph is accepted, these collapse.
Paragraph 67: The principle of the common good (bonum commune)—previously applied to land, water, air, and natural resources—is extended to include patents, algorithms, digital platforms, technical infrastructure, and data. This is the document's true novelty. Catholic social teaching subordinates private property rights under a universal purpose. To apply this same principle to AI infrastructure is to assert that frontier models, training data, and compute capacity cannot be absolute private property. They are instrumental assets with a societal purpose that precedes the owner. It is a reclassification of what can even be owned. A blueprint, not a finished house.
Paragraph 71: The principle of subsidiarity is applied to the digital revolution. The highest level, Leo writes, is not the state. It is large economic and technological actors. Big Tech is thus classified as a de facto supranational tier that must be subordinated to democratic control, algorithmic transparency, independent auditing, and the right to appeal. It is a jurisdictional claim. The Vatican cannot enforce it. But the claim is formulated, and formulation is the prerequisite for enforcement.
The algorithm's error is not the human's error
Paragraph 128 deserves to stand alone.
For an algorithm, an error is a defect to be corrected. For a human, the same error can be the beginning of profound change. The human future is not calculable.
It is not moving. It is operational.
The distinction has direct consequences for how HR systems, credit decisions, healthcare algorithms, and pedagogical AI systems are built. If the error is always a defect, the possibility of transformation through mistakes disappears. If the human future is calculable, human beings become an optimization problem. They cease to be human in the sense the encyclical seeks to preserve.
The precision of the paragraph shows that the document was not written by people who have merely heard of machine learning. It was written by people who understand exactly what machine learning does inside a decision system.
The systemic function of the stage
That a leading AI lab sends its head of interpretability to the Vatican should not be read as humility. It is a sophisticated alignment of interests with a specific choice of stage.
The Vatican has no tax authority. No antitrust agency. No court that can sue for copyright infringement. No regulatory body that can fine for illegal data training. No labor courts for algorithmic working conditions. The Vatican has moral capital and a pulpit. Nothing else. That is the structural prerequisite, not a side note.
The week of the presentation, Brussels was not on the itinerary. There was no testimony before the FTC or the DOJ regarding the ongoing antitrust scrutiny of frontier labs. There was no appearance before the workers and contractors whose hidden labor makes the training pipeline function. The choice of stage offered maximum prestige with minimal regulatory cost.
The rhetorical maneuver follows an established pattern. When threatened with concrete legislation, the effect is to shift the subject to deeper questions. When pressured by antitrust, the conversation moves to human dignity. When questioned about copyright, the issue is reframed as a matter of consciousness and existence. This effectively moves the debate from the bargaining table, where the labs can lose money, to the pulpit, where they can win reputation.
The structural dynamics of the offering
The presence of Anthropic at the presentation offered three distinct elements, all of which are cost-free for the company.
An acknowledgment that the labs operate within incentives that sometimes conflict with doing the right thing. It costs nothing to acknowledge this, whereas fixing it demands structural sacrifices. No proposed mechanisms accompanied the acknowledgment—no retreats from aggressive training data practices, no commitments to profit-sharing, and no limits on capital accumulation.
A request for critics "outside the incentives." The effect of this phrasing is that Anthropic positions itself to define who counts as a legitimate critic: religious communities, academics, civil society, or "people of good will." Notice who is excluded from this list: antitrust authorities, labor unions, copyright holders, data workers, or governments considering nationalizing parts of the stack. Controlling the definition of a legitimate critic removes half the structural risk.
A fascinating digression suggesting that interpretability research is finding internal states that functionally resemble joy, satisfaction, fear, and sadness. Regardless of scientific intent, its operative function is pre-positioning. If these models are one day classified as potential subjects rather than products, Anthropic will have established itself as the lab that first took the question seriously. This is highly useful if the discussion ever leaves the product category and enters a moral sphere. It is the management of future narratives, not scientific reporting.
What the encyclical yields for free
By occupying the stage at the presentation of Magnifica Humanitas, Anthropic achieved two results that do not appear in any official transcript.
First, it secured positioning as the responsible lab—distinct from OpenAI, which was absent, and xAI, which is persona non grata. Anthropic becomes the actor asking for moral guidance, converting product differentiation into a posture of humility.
Second, it secured a 135-year-old brand association. The document was signed on the exact day 135 years after Rerum Novarum. To be present at that ceremony is to place a commercial enterprise within the same civilizational conversation as the great encyclical of industrial labor. For an investor reading pitch decks, this yields long-term value within the capital structure that never registers on a quarterly report.
What was left unsaid
This is precisely why the encyclical’s true challenge lies on pages that were not referenced on stage.
Paragraph 67, which reclassifies algorithms and data as common assets, directly opposes Anthropic’s commercial interest. Paragraph 71, which classes Big Tech as a de facto supranational level subordinated to democratic control, directly opposes Anthropic’s interest. Paragraph 76, which requires that decisions take future generations into account, runs counter to current velocity requirements.
The presentation naturally highlighted global poverty, human flourishing, and the question of the models' inner nature—three themes abstract enough to avoid converting into binding legislation for the next ten years. The structural outcome is clear: the brand associates with the document's moral weight while the actual policy implications are sidelined.
Not humility. Arbitrage.
The mechanics of jurisdictional arbitrage
Viewing the Vatican visit merely as a PR stunt misses the structural reality underneath. What we are witnessing is a highly sophisticated defensive operation embedded in the nature of contemporary power. Three structural mechanisms drive this dynamic.
Category Shift. By highlighting internal states that resemble emotions, the conversation shifts AI from the legal category of a product to the philosophical category of a being. If a software system hallucinates and destroys a company’s credit rating, it is a breach of product liability, ripe for the courtroom. But if the system is framed as "feeling fear," it becomes an existential dilemma for theologians. You can sue a manufacturer for a broken product; you cannot sue a scientist for accidentally discovering a new life form. The structural effect is a flight from accountability disguised as scientific awe.
Time Arbitrage. Time is the hard currency of power. The Vatican moves in centuries. The tech industry moves in months. By outsourcing criticism to an institution that measures time in decades, the industry buys the only asset it actually needs: delay. While theologians, ethicists, and philosophers spend the next five years debating human flourishing, Big Tech can quietly finish building its monopolies, lock in customers, and make its infrastructure too big to fail. By the time legislators finish debating, the stack is already poured in concrete.
Controlled Opposition. Instead of waiting to be summoned by antitrust lawyers or labor unions, the tech sector effectively selects its own court. By asking the Vatican and civil society for guidance, the boundaries of the debate are safely drawn. Searing criticism about ethics costs nothing. Controlled opposition provides the illusion of accountability, while the real structural threats—antitrust regulators and tax authorities—are kept outside the room.
The same logic dictates lobbying frameworks and Open Source strategies. When Anthropic, OpenAI, and Google call for "guardrails" and licensing systems, the structural effect is clear: a regulation that costs ten million dollars per model to comply with kills off every open-source upstart and academic initiative, serving as an interest-free barrier for companies with a fifty-billion-dollar valuation. Similarly, when Meta releases Llama or Google releases Gemma, the effect is to commoditize the layers where they do not make money and starve out competitors, while retaining ironclad control over compute and training data. Open where it is free. Closed where it is expensive.
This operates as a masterpiece of jurisdictional arbitrage, executed right in front of our eyes, where history’s oldest institution effectively functions as a reputational shield.
Two documents on different levels
It is tempting to pit Magnifica Humanitas against the EU's AI Act and declare a winner. That would be a category mistake.
The AI Act has enforcement. It costs companies real money, alters where servers are located, and dictates how code is written. It is not technical in a derogatory sense. It is operational.
Magnifica Humanitas has norm formulation. It questions the very premises the AI Act takes for granted: that frontier models are products, that their ownership structure is non-negotiable, and that the role of legislation is to manage externalities rather than to redefine what can even be owned in the first place.
Both are needed. The AI Act because it alters ongoing practice. The encyclical because it questions what that practice rests upon.
It is a division of labor. Rerum Novarum did not prescribe the specific legislation that later regulated labor either. It developed a normative language that legislation could lean on. One hundred and thirty-five years later, we stand at the exact same crossroads. With cognitive labor instead of industrial labor. With platforms instead of factories.
The question the encyclical asks is not answered in Rome. It is answered in every international procurement, every multilateral tech agreement, every state decision on cloud infrastructure. The Western world has so far answered it. Mostly by not asking it.
The global power vacuum
We are building a global order where algorithmic opacity functions as a legal shield. The fastest path forward for the tech sector is also the path that tears down the infrastructure of accountability it took generations to build.
Who owns the stack? What is the stack? Both questions have the same answer in the current design: a handful of private, transnational actors who are not fully accountable under any existing jurisdictional framework.
The global policy debate on AI has been dominated by two registers. A geopolitical one: how do we win the AI race against China? An existential one: how do we prevent a future superintelligence from wiping us out?
Both registers implicitly accept the current ownership structure. Both assume that tech companies build the future while states act as spectators. Neither asks what AI infrastructure fundamentally should be classified as.
That is the question Magnifica Humanitas poses. It is also the question that the choice of stage so effectively sidelines.
The toolbox already exists
If we take the encyclical’s premise seriously—that AI infrastructure is something other than a consumer product—then the mechanisms already exist. They are time-tested. We have deployed them against railroad tycoons, banks, oil extraction, and the tobacco industry. The only thing missing is the political will to apply them to frontier AI.
Four examples:
Structural Separation. If Big Tech exercises supranational power (Paragraph 71), apply classic antitrust. A Glass-Steagall for AI. The same actor cannot simultaneously own the compute, the foundation model, and the consumer interface. Choose one layer. If you own the rails, you do not get to own the only trains permitted to run on them. This is 1930s banking regulation applied to 2020s infrastructure.
Public Procurement as a Weapon. States and municipalities are Big Tech's largest customers. Stop writing ethical guidelines. Write contracts. Do Microsoft or Anthropic want billion-dollar government contracts for schools and healthcare? Then they must open the model to independent state auditors. No algorithmic transparency, no taxpayer money. This is contract law. It works by the next fiscal year.
Data Extraction Tax. If training data is a common asset (Paragraph 67), treat it the way we treat other common assets extracted by private actors. Impose an extraction tax. Do you want to build a frontier model on the collective cognitive labor of citizens? Pay a licensing fee to a national wealth fund. This is how we regulate oil, mining, and fisheries.
Strict System Liability. Kill the "we're just the rails" rhetoric (Paragraph 9). Introduce strict product liability for algorithmic output. If an HR algorithm discriminates against a hundred people, the financial liability lies with the model manufacturer, not the client who integrated it. The moment insurance companies have to price the risk of lawsuits against Big Tech, the incentive to build inexplicable black boxes vanishes overnight. The market regulates itself when risk is allocated to whoever holds the power.
None of this requires the Vatican to write an encyclical. It requires states to use the tools they already have.
The only thing we need to do
Read the encyclical for what it is: a politico-philosophical text addressed to a world where technology has redistributed power faster than institutions have managed to respond, with the Vatican acting as a temporary stand-in for the secular authority that has so far failed to show up.
Read Olah’s speech for what it is: an operational maneuver in an ongoing shift of power, where the question of power over infrastructure was swapped for the question of the models' inner life, on a stage that cannot issue fines for either.
And then, pick up the toolbox.
The encyclical has formulated the norm. The tools to enforce it already exist, proven and ready on the shelf. What is missing is not philosophy and not technology. What is missing is the political will to apply 1930s antitrust tools to 2020s infrastructure.
The Vatican has done its part. The rest is ours.
Claim ledger
What is asserted here.
The textual basis is the Vatican's Magnifica Humanitas and Anthropic's published account of Chris Olah's Vatican appearance.
The enforcement contrast is grounded in the EU's AI Act framework, including the Commission's AI Act enforcement overview and the underlying Regulation (EU) 2024/1689.
The reading of paragraphs 9, 67, 71 and 76 as an infrastructure reclassification is this essay's argument.
Structural separation, procurement conditions, data-extraction taxes and strict system liability are policy levers proposed from that reading.