I read Pope Leo XIV’s first encyclical the way you would read an audit of your own company. Not as a religious text. I read it as a practitioner.
On 25 May, the Pope presented his first encyclical in person, a break with tradition, and stood next to Chris Olah, co-founder of Anthropic. Olah told the room that every frontier AI lab, including his own, operates inside incentives that can pull against doing the right thing. He said the field needs people outside those incentives. People willing to say hard things. That is a co-founder of one of the major labs publicly endorsing the pope’s message.
Most coverage missed it. The press settled on three angles. The Pope warns about AI. The Pope takes aim at big tech. The Pope makes AI ethics a religious matter. All three are true. None of them is useful if you build, sell, or buy this technology for a living.
If you read as an audit, three things stand out. Each one names what the industry already knows and rarely admits.
One. The design problem.
Paragraph 98 says current AI systems are more cultivated than built. Even the people who build them have a limited understanding of how they work. Internal representations and computational processes, as the text puts it, remain unknown.
That is what the interpretability literature already says about itself. A 2024 review concludes that complete understanding of frontier systems should not be expected. A 2026 paper documents a knowledge to action gap. We can sometimes see what is happening inside a model and still cannot correct what it does.
The Pope did not invent that critique. He read what the field publishes about itself and put it in plain language.
For a practitioner, that costs something. When a customer asks how a model arrived at a decision, the honest answer is closer to “we cultivated it” than “we engineered it.” The regulator and the enterprise buyer are about to start expecting that answer.
Two. The governance problem.
Paragraph 107 says a more moral AI is not enough if that morality is determined by a few. Alignment, run as a closed project inside the labs themselves, does not stand up to the question of who decided.
Olah said the same thing in different words. External oversight. Outside voices. People willing to say hard things. From the co-founder of one of the largest labs in the world.
The EU AI Act becomes fully applicable on 2 August 2026. Three months from now. High risk AI systems will require third party conformity assessment before deployment. General purpose models with systemic risk already face disclosure obligations on training data.
The Pope, the AI lab co-founder, and the regulator are saying the same thing: External oversight is not a religious request. It is the operating mode for save AI deployment whether we want it or not. For us as practitioners, that is not a debate to win. It is a deadline.
Three. The supply chain problem.
Paragraph 173 says nothing in the world of AI is immaterial or magical. Every fast clean answer rests on a chain of work. Data labelers. Content moderators. Miners of rare earths. Often young, often women, often working for very little.
The numbers are public. Kenyan data labelers report earning about two dollars an hour. Comparable workers in the United States earn more than twenty. In February 2025, Kenyan labelers formed the Data Labelers Association, the first organized labor body for this workforce. Lawsuits against Meta over moderator working conditions are active in Spain, Kenya, and Ghana.
The industry treats this as a PR problem. The Pope treats it as what decides whether the technology is legitimate at all. A due diligence reviewer would flag it in the first hour of a data room. We can fix this. Yet we can also ignore it and watch a customer or investor read paragraph 173 back to you with a red pen in hand.
One counterargument is worth naming. Slowing down kills the industry. Disarming AI is a euphemism for losing to China. The encyclical does not say stop. Paragraph 110 says to disarm does not mean rejecting technology, but preventing it from dominating humanity, and freeing it from monopolistic control. The competitive position to worry about is the one built on a labor chain or a governance model that does not survive the next regulator reading the next paragraph.
So what to do? Three thoughts:
When you describe how your model works, tell the truth about what is cultivated and what is engineered. The buyers and regulators who matter will respect honesty more than another launch deck.
Treat external governance as a product feature, not a tax. Build for the third party conformity assessment that lands 2 August. The labs and vendors that arrive early will set the standard.
Make your supply chain visible. Audit data labeling, model training, and content moderation. Pay it like the work it is. Publish what you find. This is the part of the job that turns into a lawsuit if you wait.
The Pope read our industry better than most of our boards would put on paper. The mistake is to read him back as a sermon. Read him as a good knowledgeable auditor.
