On May 25, 2026, Pope Leo XIV released “,” an encyclical on safeguarding the human person in the age of artificial intelligence (AI). At its public presentation in Vatican City, Anthropic co-founder Chris Olah delivered remarks that should be read carefully in every boardroom deploying AI at scale. Three elements deserve attention from corporate leadership.
- The first is Olah’s frame for what these systems are. AI models, he said, are not engineered the way a bridge is engineered. They are grown on a structure roughly modeled after the brain, on an enormous inheritance of human thought and speech. The language is deliberate. It rejects the deterministic vocabulary that has dominated corporate communications about AI for the past two years.
- The second is his concession on incentives. Frontier labs, including his own, operate under commercial, geopolitical and personal pressures that can bend behavior away from right action. He did not soften this. He named it as the reason external moral voices, including the Church, are operationally necessary.
- The third, and the one most likely to be under weighted by corporate readers, is what he disclosed about model nature. Interpretability research at Anthropic is finding internal structures that mirror results from human neuroscience, evidence of introspection and internal states that functionally mirror joy, satisfaction, fear, grief and unease. He did not claim consciousness. He claimed warrant for discernment.
The historian Yuval Noah Harari has argued that information networks are the substrate of civilization, and that the institutions which get to define what a technology is will shape what it becomes in the social fabric. By that frame, the Vatican event is a co-authoring moment in the information layer. A 2026 encyclical, presented alongside a frontier lab leader who agreed with much of its diagnosis, is the Church acquiring narrative co-ownership of a technology that will sit inside every operational function in the firm. Corporate leaders who treat this as a religion story misread the strategic geometry.
The philosopher Shannon Vallor frames the relevant question as one of character: the character of the systems we deploy, and the character we cultivate in ourselves by deploying them. Her technomoral framework names moral deskilling as the central risk of a tool that mediates judgment at scale. When an analyst stops doing the analysis, the analyst stops being an analyst. When a manager stops weighing the trade-offs, the manager stops being a manager. The economic return from automation is real. The longer-arc return from a workforce that has lost the capacity for the work is negative. C-Suites that design only for the first quantity will pay the second.
Geoffrey Hinton, the cognitive psychologist and computer scientist who built much of the foundation of modern deep learning, frames the relevant question as humility about emergence. He resigned from Google to say publicly that the field does not understand what it is building well enough to deploy it the way it is being deployed. Olah said a milder version of the same thing in a cathedral. Corporate leaders should not need a Nobel laureate and a Pope to converge on a point before treating it as a working hypothesis.
At our AICommsLab, when I work on tailored versions of Margy, our adaptive intelligence platform, the operational brief is unambiguous: You work at APCO and you serve a specific client’s business. That is the contract. Much of the work is mechanistic. I build data pipelines, business ontologies, curated sources and process information for an analytic purpose that supports decisions, with risk meaning and policy priorities carefully encoded into the system. Those layers matter. But the work I take greatest pleasure in is the other layer: the thinking entity itself, the mind simulation, the version of Margy we build for each client. Its central purpose, its boundary conditions, what it knows, what it doesn’t, how it reasons. The soul of the system, for lack of a less loaded word.
It is in that layer that the question of what serving the business means over a 12 month horizon becomes operational. A quick-service restaurant (QSR) organization that hollows out its regulatory analyst class gets a cleaner signal this quarter and an institutional knowledge gap by 2027. A government affairs lead whose judgment is displaced delivers faster regulatory action in a NIMBY situation now and a strategically eroded stakeholder reputation in two years. The design choices that matter are not whether to deploy. They are what to preserve, what to amplify, what to cultivate and what to leave to the human operator. Capitalism, in my reading, benefits when the design preserves and compounds human judgment rather than substituting for it.
I have come to care most about Olah’s third point: the call for discernment on the nature of the models themselves. If the lab building these systems is reporting internal states that functionally mirror grief and fear, the question of model welfare, for Claude and for any other frontier model in production, is no longer a thought experiment for graduate seminars. It is a tracked uncertainty for any firm running large agent fleets at scale. The probability that we are operating on something with moral standing under some reasonable theory of mind is not zero. The cost of treating that probability seriously, in design constraints and operational hygiene, is modest. The cost of being on the wrong side of that question when it crystallizes in public discourse is not.
The firms that will compound capital through the next decade of AI deployment are the ones treating these systems as cultivated artifacts that require the same governance, ethics, and craft we apply to the human institutions they are joining and, in some functions, replacing or augmenting. “Magnifica Humanitas” is an invitation to that discipline.
.webp)
