AI ethics and the responsibility behind the model

The Pope wrote 245 paragraphs on AI. If you work in this industry, read it before you decide it does not apply to you.

Magnifica Humanitas is the first encyclical of Leo XIV’s papacy, signed on the 135th anniversary of the document that defined modern Catholic teaching on industrial labor. He is telling you he thinks this is that kind of moment.

The convenient dismissal is that he is a religious leader and the tech world is famously allergic to that. The harder part to ignore is that he speaks for more than 1.4 billion people. The Holy See has been talking directly to Silicon Valley for nearly a decade through the Minerva Dialogues, and on the stage at the launch was Chris Olah, co-founder of Anthropic, telling the room the labs need “moral voices that the incentives cannot bend.” Not a script you would expect from one of the people actually building frontier models.

Whatever you think of the Vatican, the people drafting these documents are not uneducated. Strip the religious vocabulary off. Here is what they are saying.

Technology is never neutral. Every algorithm carries the values, blind spots, and commercial incentives of whoever built and funded it. In security we have been having this conversation about software supply chain for years. Training data is the supply chain. Fine tuning is the supply chain. The reinforcement learning from human feedback rater pool is the supply chain. Almost none of it can be independently scrutinized today, and the labs are not in a hurry to fix that.

AI is rapidly becoming infrastructure disguised as product. Five vendors, give or take, sit underneath nearly every meaningful model, on top of an infrastructure layer of minerals, water, and power drawn disproportionately from the Global South. “AI can very easily be colonial,” one panellist said. That phrase will age well.

Consequential decisions should not be handed to systems that do not know compassion, mercy, or forgiveness. Credit scoring. Hiring. Criminal risk assessment. The apparent objectivity of a score launders the choices designers made upstream, and the person on the receiving end has no real route to challenge it.

Autonomous weapons are a line. Leo declares the old “just war theory” outdated and says handing the decision to kill to a machine is a moral failure, not a technical achievement. The public conversation in defence and intel has been some variation of “the other side is doing it, so we must.”

That is not an argument. It is a shrug.

Leo is refusing the shrug.

A more moral AI is not enough if the morality is set by whoever controls the data and the compute. That line should make every CEO in this industry uncomfortable.

You can disagree with the messenger. You will struggle to disagree with the argument.

The Vatican has done its homework.

Most of this industry has yet to do its own.