A Full Analysis of Pope Leo XIV's AI Encyclical: How the Vatican Views Artificial Intelligence Ethics

On May 15, 2025, the Vatican released Pope Leo XIV's encyclical Magnifica Humanitas (The Light of Humanity), comprehensively articulating how to protect human dignity in the age of artificial intelligence. Tech commentator Simon Willison called it "one of the clearest texts I've ever seen on the ethical issues of AI integration into modern society." This technical ethics document from a religious leader offers a unique and profound perspective on the current cacophonous AI debate.

From the Industrial Revolution to the AI Revolution: Echoes of History

Leo XIV's choice of the papal name "Leo" was no coincidence. He explicitly pays tribute to Leo XIII, who issued the encyclical Rerum novarum (Of New Things) in 1891—a document that systematically articulated the rights and duties of capital and labor in the context of the First Industrial Revolution, profoundly influencing social thought for more than a century thereafter.

To appreciate the weight of this tribute, one must revisit the historical context of Rerum novarum. Europe in 1891 was in the throes of violent upheaval from the Industrial Revolution: the rise of the factory system, the mass formation of an urban proletariat, the prevalence of child labor and appalling working conditions, and the collapse of the traditional feudal order and guild system. Rerum novarum was the first systematic articulation of workers' dignity and rights, the legitimacy and limits of private property, the boundaries of state intervention in the economy, and the mutual obligations of employers and employees. This document inaugurated the tradition of "Catholic Social Teaching," which successive popes developed through a series of encyclicals, forming a complete ethical system encompassing labor, property, development, and justice. By choosing the same name to echo this tradition, Leo XIV declares that the AI revolution's impact on human social order is of the same historical magnitude as the Industrial Revolution, demanding an equally serious moral response.

In his first formal meeting with the College of Cardinals, Leo XIV explained: "In our time, the Church offers the treasury of its social teaching to all people, in response to another industrial revolution and the new challenges brought by developments in the field of artificial intelligence—challenges concerning the defense of human dignity, justice, and labor."

This historical analogy itself conveys an important signal: the Vatican views the AI revolution as a social transformation of the same magnitude as the Industrial Revolution, requiring an equally serious ethical response.

The Explainability Problem of LLMs: "Grown" Rather Than "Built"

Paragraph 98 of the encyclical offers an exquisitely precise description of the nature of large language models:

Current AI systems are less "built" than they are "grown," because developers do not directly design every detail but instead create a framework within which intelligence "grows." Consequently, fundamental scientific aspects such as the internal representations and computational processes of these systems remain currently unknown.

This metaphor is precise because it captures the technical essence of modern AI development. Large language models are not created by engineers writing logical rules line by line; rather, through gradient descent optimization on massive text datasets, tens or even hundreds of billions of parameters are automatically adjusted to a state where they can predict the next token. The result of this process is a numerical matrix in a high-dimensional parameter space—with no human-readable rules, no explicit knowledge representation, only compressed encodings of statistical correlations. Even the development teams behind GPT-4 or Claude cannot directly "look up" why the model gives a particular answer to a specific question. Companies like Anthropic have established dedicated "Interpretability Research" teams that attempt to reverse-engineer internal model representations through activation analysis, feature attribution, and other methods, but this work remains in its early stages. The encyclical's use of the word "grown" corresponds precisely to the machine learning concept of "emergence"—complex capabilities are not designed in but arise spontaneously during training, in ways that even researchers cannot predict.

This "grown vs. built"