Magnifica Humanitas and the Work of Building Good Tools
On May 25, 2026, the Vatican released Pope Leo XIV's first encyclical, Magnifica Humanitas. It runs to five chapters and 245 numbered paragraphs, and underneath all of them is one question: what happens to human dignity when we hand our thinking to machines? Leo signed it on May 15, the 135th anniversary of Rerum Novarum, the letter Leo XIII wrote when the last industrial revolution was breaking families and rewriting what work meant. The date is the argument. This Pope reads machine intelligence as another turn of the same wheel.
The claim at the center is simple and not soft: technology is never neutral. It takes on the character of the people who design it and the institutions that fund and regulate it. Treat a tool as innocent and you lose the ability to see how it quietly reshapes the world, and your own agency along with it, toward surveillance, concentrated power, and the habit of treating people as data.
The serious voices were all saying it
It is not only the Vatican raising the alarm. The most serious enterprise operators are seeing the same pattern in a different vocabulary.
Lori Beer, the global CIO of JPMorgan Chase, has been blunt that responsibility and accountability cannot be outsourced at the point where business-critical decisions are made. Marco Argenti of Goldman Sachs warns of a gigawatt ceiling, a near future in which a handful of hyperscalers control the computing power required to scale and, with it, who gets to scale at all. Read side by side, the Pope and the two CIOs are circling one anxiety: intelligence without governance is not infrastructure. It is exposure.
The missing middle is practice
Most of what we read about AI is either breathless promise or fatalistic alarm. Both are easy to write, and neither helps the person building a product, writing a book, or running a team. What is missing is the practitioner's register: an honest account of what it takes to use these tools well, which begins with the fact that the design of a tool decides which way it tips. For the engineer shipping a workflow or the writer finishing a manuscript, using AI well means deciding which steps stay governed, which steps are cognitive, and who signs off when the machine is wrong.
I feel the weight of that every week now, in calls from people whose work has been declared surplus overnight. The machine has made their artifact cheap, the draft, the deck, the model. But the value in that work was never the producing. It was the judgment they exercised before their name went on it, and judgment is the part that survives. To build well is to build so that judgment stays where it belongs.
My own work as a laboratory
For the past eighteen months I have treated my own work as a laboratory for this tension. Across the Cost of the Machine trilogy, the building of OakQuant Press, and the tools around it, I have been testing one hypothesis: that human judgment is a kind of friction worth protecting rather than smoothing away.
I use AI as a partner for drafting, refactoring, and testing, but I guard the seams of the work, the framing and the structure and the final judgment, with extreme care. Speed is a side effect. The point is continuity of thought across domains. When I build, I am building bounded agents that extend my intent rather than replace it.
The cognitive budget
In my professional work I help large enterprises build toward what I call the accountable agentic enterprise, where the goal is not only efficiency but auditability: the ability to say who is responsible when an agent acts.
The discipline that makes that possible starts small. Treat model intelligence as free and infinite and you slowly hand the work over, one easy step at a time. So I keep a cognitive budget, spending model cognition only where judgment is the actual work and routing everything else through deterministic, governed coordination.
The cognitive cost of running a process at scale is roughly , where is the volume of cases, is the percentage of steps routed to model cognition, is the tokens each cognitive step consumes, and is the price per token. This formula is not an exercise in optimization; it is a budget for your own agency. Keep that percentage as low as you can, reserving model cognition for the steps where judgment is genuinely the job, and you have kept your judgment. Let it trend toward 100 percent and you are no longer directing the tool. You are slowly becoming its assistant.
Designing for dual use
These tools are not harmless, and pretending otherwise would cost the argument its credibility. Any system that can synthesize information can also be pointed at flooding the world with synthetic certainty, targeted persuasion, or automated surveillance.
The aim of a good tool is not to deny dual use but to build with it assumed: clear boundaries, audit logs, and human sign-off where it counts, with a standing refusal to confuse machine fluency with authority.
What it means to build accordingly
The encyclical's question is not whether AI is good or bad. It is whether AI stays answerable to the human person. Answering it is not the work of agreeing with a Pope. It is the work of building as though the answer were yes: tools with seams, decisions with sign-off, and judgment kept rather than surrendered every time a machine hands you something plausible.
Leo frames the whole letter between two builders. There are the men at Babel, who pooled their power to reach the sky and lost the ability to understand one another. And there is Nehemiah, who rebuilt the walls of his city stone by stone, answerable to the people who would live inside them. The encyclical is really asking which kind of building this is.
Technology is never neutral. It takes on the character of whoever shapes it. The quality of our tools will be measured not only by what they can do, but by what they make it easier for human beings to remain.