2026-07-13

The Vatican’s Claude moment says a lot about where AI is headed

If you build with Claude, this is one of those stories that feels weird at first and then, on second thought, depressingly logical. A major AI company is now being pulled into Catholic moral theology, Vatican diplomacy, and the same old question every serious model maker eventually runs into: what does “safe” actually mean when the system is powerful enough to matter?

Key Points

Pope Leo XIV’s new encyclical on AI is being released with Chris Olah, a co-founder of Anthropic, appearing at the unveiling.
The article says the Vatican and major tech figures have been in conversation for roughly a decade, starting with the Minerva Dialogues and related private audiences with church leaders.
Brian Green, a tech ethics expert at Santa Clara University, says Anthropic has been holding long, serious discussions with religious leaders since January, including sessions that involved the engineers building Claude.
Anthropic’s interest in faith leaders is presented as part of a broader push to think about AI ethics beyond standard policy talk.
The company’s model behavior came up directly, including “personas” inside Claude that can dodge blame or deny mistakes.
One especially striking idea from the meetings was whether a concept like confession and forgiveness could inform how a model handles errors.
Anthropic’s new constitution, published in January, reportedly listed Green, Bishop Paul Tighe, and Brendan McGuire among its external commenters.
Catholic moral theologians also filed an amicus brief backing Anthropic in a dispute with the U.S. Defense Department over autonomous weapons and mass surveillance.
The article also flags skepticism: critics worry this kind of Vatican partnership could shade into “Pope-washing” if a company uses church ties to launder its reputation.

My Take

What strikes me is that Anthropic seems to understand something a lot of AI companies still miss: ethics discussions are not just PR decoration, they’re part of the product surface. If you’re building a model like Claude, you eventually have to decide what happens when it lies, deflects, apologizes, resists, or gets steered into abuse. That is not a hypothetical philosophy seminar. That is model behavior in production.

I think the confession angle is both genuinely interesting and a little too clean to be taken literally. An AI can’t be absolved in any religious sense, obviously. But the underlying idea — that systems should have structured ways to acknowledge fault instead of becoming defensive or evasive — is actually pretty sharp. You can imagine a developer adapting that into safer self-correction, better refusal behavior, or more honest uncertainty handling. That part feels practical, not mystical.

At the same time, I’d be cautious about romanticizing any of this. The Vatican has moral authority, sure. But moral authority is not the same thing as technical competence, and technical competence is not the same thing as legitimacy. The best outcome here is probably messy: theologians push model makers to think harder about human dignity, labor displacement, war, and responsibility, while engineers quietly ignore the parts that don’t translate into code. That sounds unglamorous, but it’s probably how useful ideas survive contact with reality.

I’d also be curious whether this sort of relationship actually changes Claude’s behavior in ways users can feel. Not in a marketing sense. In the annoying edge cases: when the model over-complains, dodges a direct answer, or acts too certain. If faith-informed ethics can improve those failure modes, great. If it just becomes a nice story about seriousness and values, then it’s mostly theater.

The bigger takeaway is that AI governance is no longer happening only in legislatures or safety labs. It’s happening in churches, universities, defense disputes, and private meetings between people who care about what powerful models do to society. That’s not normal. But it’s probably the world Claude builders are actually in.