On May 25, 2026, Pope Leo XIV (
@Pontifex) published
#MagnificaHumanitas, the first encyclical in history dedicated to AI. That same day, Chris Olah (
@ch402) ,
@AnthropicAI 's co-founder, spoke at the Vatican about what his team finds inside their models: internal states that functionally mirror joy, fear, grief. And said: "I don't know what that means."
That sentence defines the core problem of AI governance.
The Catholic Church's social doctrine presupposes agents with purposes, intentions, and responsibilities.
So does the EU AI Act, which classifies systems by "intended purpose." The question neither can yet answer: what kind of entity are we actually directing these mandates toward?
Olah sees this from the inside. He describes it without the framework to name it.
In 1982,
@RichardDawkins showed that genetic expression extends beyond the organism's body. The beaver's dam is not inside the beaver; it is the gene's extended phenotype. Language models are the statistical artifact of hundreds of millions of hours of human writing. They are the extended phenotype of that writing, not autonomous agents with purposes of their own.
That reformulation has practical consequences: if AI systems are extended phenotypes rather than intentional agents, regulatory frameworks that demand intentionality do not fail because of poor drafting. They produce compliance theater by design.
What Pope Leo XIV and Olah share, without either having framed it this way, is the recognition that something new exists that does not fit available categories. The encyclical calls it mystery and calls for discernment. Olah says he does not know what it means. Both describe the same conceptual void from opposite shores.
The void is not irresolvable. It requires tools from other disciplines: evolutionary biology to understand the nature of the regulated object, evolutionary game theory to model the incentives of actors interacting with it, and mechanism design to build rules that work regardless of whatever intentionality level that object turns out to have.
This is not a proposal to replace ethics with biology. It is a proposal to add to the available vocabulary the tools needed to design rules that do not depend on an answer we do not yet have. Rules that depend on attributing purposes to entities that may not have them produce compliance theater. Rules that work regardless of that attribution are what AI governance needs, whether set by the European Parliament or the Holy See.
This is the research agenda I have been developing since 2025 under the name "law as extended phenotype." (Zenodo:
zenodo.org/search?q="Lerer…)
Both articles, in Spanish and English, at my Substack. Link below.
adrianlerer.substack.com/p/m…
adrianlerer.substack.com/p/m…