The Vatican Admission

On May 25, 2026, Christopher Olah walked to a podium in the Vatican's Synod Hall and stood next to Pope Leo XIV. Olah is a co-founder of Anthropic, the company that sells itself as the safety-first AI lab. He was there to co-present the pope's first encyclical, Magnifica Humanitas, a 20,000-word condemnation of the industry Olah helped build. Then he told the room that AI development "operates inside a set of incentives and constraints that can sometimes conflict with doing the right thing."

Read that sentence again. A co-founder of a company valued at over $800 billion, a company that just posted its first quarterly profit on $10.9 billion in Q2 revenue, went to the seat of the Catholic Church and admitted the incentive structure of his own industry conflicts with doing the right thing.

This is not a leak. It is not a disgruntled employee. It is not a whistleblower. It is a co-founder, standing in an official capacity, at an event his company chose to participate in, co-presenting a document that accuses his industry of threatening democracy, widening inequality, and racing toward autonomous weapons. He agreed with the document. He helped present it. And he added his own admission on top.

The question is not whether Olah was sincere. He probably was. The question is what it means when the person responsible for understanding how these systems work tells you, on a global stage, that the financial machinery around those systems pushes against safety.

The Scene

Pope Leo XIV personally presented Magnifica Humanitas at the Synod Hall on May 25. This was unusual. Popes sign encyclicals. They do not typically stand at a podium and walk audiences through them. Leo XIV did.¹ The document had been signed ten days earlier, on May 15, timed to the 135th anniversary of Rerum Novarum, the 1891 encyclical by Leo XIII that defined Catholic social teaching on labor and capital. The parallel was intentional. Leo XIII wrote about the exploitation of industrial workers. Leo XIV was writing about the exploitation of everyone else.²

The encyclical frames artificial intelligence as "not a technological challenge, but an anthropological one."³ That framing matters. Technology challenges have engineering solutions. Anthropological challenges require asking who benefits, who is harmed, and what kind of society the technology creates. The Vatican was explicitly rejecting the Silicon Valley framing in which AI is a tool to be optimized. Instead: AI is a force reshaping what it means to be human, and the people building it are not asking the right questions about that reshaping.

Magnifica Humanitas runs long and covers wide ground. It condemns lethal autonomous weapons. It warns that Big Tech companies risk widening inequality and weakening democratic institutions. It addresses worker displacement. It calls out the simulation of human faces and voices without consent. And it makes an appeal that would sound naive coming from anyone other than the head of a 1.4-billion-member institution: "disarm AI" from military and economic interests.³

Sitting in the audience and then standing at the podium was Christopher Olah.

The Man at the Podium

Olah is not Anthropic's CEO. That is Dario Amodei. Olah is the co-founder who leads interpretability research, the branch of AI safety that tries to understand what is happening inside neural networks. If Anthropic's public identity is "the safety lab," Olah is the person who does the actual safety science. He is the one opening up model internals to see which circuits fire, which features activate, which patterns the models are learning that their creators did not intend.

Anthropic chose to send him. That choice is worth sitting with. They did not send a policy person. They did not send a communications executive. They sent the researcher whose work is most directly about understanding the gap between what AI systems are supposed to do and what they actually do.

And that researcher, standing in the Vatican's Synod Hall, said something that the quarterly earnings calls and IPO prospectuses will not say: the incentive structure of AI development can conflict with doing the right thing.

"Operates inside a set of incentives and constraints that can sometimes conflict with doing the right thing."

This is a remarkable sentence to say out loud, in public, from a podium, at an event covered by global press. Every word carries weight. "Operates inside" means the conflict is structural, not incidental. "Incentives and constraints" means both the carrots (revenue, valuation, market share) and the sticks (competition, investor expectations, the race to ship). "Can sometimes conflict" is the diplomatic version. The undiplomated version is: the business model and the safety mission are pulling in opposite directions, and the business model is winning enough of the time that a co-founder felt compelled to confess it at the Vatican.

The Numbers That Surround the Confession

Consider where Anthropic stands financially at the moment Olah said those words.

The company posted its first quarterly profit in Q2 2026, with $10.9 billion in revenue for the quarter, on an annualized run rate approaching $44 billion.⁴ Google has committed $40 billion in total investment. An IPO is expected as soon as October, with a target valuation north of $800 billion.⁵

Anthropic also announced a $200 million partnership with the Gates Foundation to deploy AI in underserved regions.⁶ That announcement landed the same week as the Vatican event. The optics are coordinated. Safety lab co-presents papal encyclical. Safety lab partners with Gates Foundation for global good. Safety lab prepares for an IPO that will make its founders billionaires several times over.

These numbers are the incentives Olah was talking about. You do not build a company to a $44 billion run rate by prioritizing safety over shipping. You do not attract $40 billion from Google by moving slowly. You do not achieve an $800 billion valuation by telling investors that your interpretability research found something concerning and you need to pause.

And yet the interpretability lead went to the Vatican and said the incentives conflict with doing the right thing. At some point you have to take the man at his word.

The Pentagon Problem

There is another piece of context that makes Olah's Vatican appearance stranger and more significant. The Pentagon has labeled Anthropic a "supply chain risk" for refusing to develop lethal autonomous weapons systems.⁷ While other AI companies compete for defense contracts, Anthropic has maintained a policy against direct weapons applications. The Department of Defense, in procurement documents, has flagged this refusal as a risk factor for military programs that depend on frontier AI capabilities.

So picture the geometry. Anthropic's co-founder is at the Vatican, co-presenting a papal encyclical that explicitly condemns lethal autonomous weapons. Back in Washington, the Pentagon considers Anthropic's refusal to build those weapons a supply chain vulnerability. The company is simultaneously too ethical for the military and too compromised for the pope.

That tension is the incentive structure Olah was describing. The same company occupies both positions at once. It refuses weapons contracts (costing it defense revenue) while racing toward an $800 billion IPO (requiring growth that comes from somewhere). The Vatican praises the refusal. The Pentagon punishes it. The market does not care either way as long as the revenue line goes up and to the right.

What the Encyclical Actually Says

Magnifica Humanitas is, by papal standards, blunt. The Vatican has been building toward this document for years, through the Rome Call for AI Ethics in 2020 and Pope Francis's interventions before his death. Leo XIV inherited that institutional momentum and sharpened it.²

The core argument: AI systems concentrate power in the hands of a small number of corporations. Those corporations are accountable to shareholders. Shareholders optimize for returns. Returns come from deployment speed, market capture, and cost reduction. None of those optimization targets include "preserve human dignity" or "strengthen democratic institutions" or "protect workers from displacement."

The encyclical does not frame this as malice. It frames it as structure. The companies are doing what companies do. The problem is that what companies do, when the product is a general-purpose intelligence system, has consequences that extend beyond normal market dynamics. A car company that cuts corners produces dangerous cars. An AI company that cuts corners produces systems that reshape information flows, labor markets, and the boundaries of what counts as real.

The document is specific where previous Vatican statements were abstract. It names worker displacement as a present-tense harm, not a hypothetical. It calls out the simulation of human faces and voices. It condemns autonomous weapons by name. And it makes the "disarm AI" appeal, asking governments and companies to remove artificial intelligence from military and coercive economic applications.³

By signing the document on the anniversary of Rerum Novarum, Leo XIV drew a direct line. In 1891, Leo XIII told industrialists that the labor market was not a natural phenomenon beyond human control but a system designed by people who could redesign it. In 2026, Leo XIV told technologists the same thing about AI. The incentive structure is not physics. It is a choice.

The Pattern We Have Been Tracking

We have been documenting these structural contradictions for months: vendors who design products to resist auditing, who build trust boundaries meant to be bypassed, who fund surveys engineered to produce favorable headlines. Olah's Vatican statement is the executive-level confirmation. The person responsible for interpretability research described, from the inside, the same dynamic we documented from the outside. The system is working as designed. The design conflicts with doing the right thing.

The Credibility Paradox

Here is what makes this event genuinely difficult to process: Olah's statement gains its power from the fact that he is a co-founder of Anthropic. If a random AI ethics researcher said "the incentive structure conflicts with safety," that is a Tuesday. If a Vatican official said it, that is institutional positioning. But when a co-founder of the company that markets itself as the safety-first lab says it, from the Vatican podium, next to the pope, that is a confession.

And confessions create a problem for the person making them. Olah now exists in a state where he has publicly acknowledged the structural tension between his company's business and its stated mission. Every future safety decision Anthropic makes will be read through that acknowledgment. Every time Anthropic ships a product faster than its safety team wanted, or weakens a guardrail to match a competitor, or accepts investment that comes with growth expectations incompatible with caution, Olah's Vatican words will be the frame.

He cannot unring that bell. The company cannot unsend him to the Vatican. The statement is on the record, delivered at a global event, co-signed by association with a papal encyclical. Anthropic's future actions will be measured against it.

That measurement might be the point. It is possible that Olah went to the Vatican precisely to create that accountability. To put something on the record that would make it harder for his own company to quietly compromise. To use the weight of a papal encyclical as a counterbalance to the weight of an $800 billion valuation. If so, it is a remarkable gambit: using external moral authority to constrain internal commercial pressure, because the internal mechanisms alone are not sufficient.

If so, that is the most alarming interpretation of all. It means the person building the guardrails does not trust the guardrails to hold.

What Happens Next

The IPO is coming. The October timeline puts it roughly five months after the Vatican event. Somewhere in the S-1 filing, in the risk factors section that nobody reads until something goes wrong, there will be language about the company's commitment to safety. That language will be written by lawyers. It will not say what Olah said.

And somewhere inside Anthropic, Christopher Olah will go back to his interpretability research. He will continue finding things that are difficult to explain and harder to control. He will operate inside the same set of incentives and constraints he described at the Vatican. Magnifica Humanitas will be studied in seminaries and ethics courses for decades. The IPO prospectus will be studied on Wall Street for quarters.

On May 25, 2026, in the Synod Hall, Christopher Olah stood next to a pope and told the truth about his industry. The industry will spend the next year proving him right.