Papal Infallibility Meets Probabilistic Guesswork

Pope Leo XIV says AI-generated art lacks the “spark of humanity.” Also true: AI-generated deepfakes have been putting words in his mouth that he never said, convincingly enough that his own church has had to publish guidance on spotting them. One man argues the machines can’t capture a human soul, and the machines are busy counterfeiting his.

That contradiction is the most interesting thing about this story, and almost nobody covering it is treating it as the main event.

The critique is sharper than the usual art-world grumbling

I review AI tools for a living. I have seen every flavor of anti-AI-art argument, and most of them are bad. “It’s not real art” is a vibe, not a position. “It’s soulless” is unfalsifiable. “Real artists suffer” is just gatekeeping with extra steps.

Leo XIV’s version is more specific, and that’s what makes it worth taking seriously. His objection is mechanical, not mystical. These models work by learning probabilistic patterns from scraped data. Given a prompt, they produce the statistically likely continuation. That is the whole trick. It is an extremely impressive trick. But the output is, by construction, a weighted average of what already exists.

He’s drawing a line between generating a plausible image and making something that came from a particular person having a particular experience. You can disagree with where he puts the line. You can’t really dispute the description of how the tool works, because that’s the actual architecture.

What the argument gets right

Reports of his position emphasize imperfection, singularity, and the irreducibility of human experience as things art should protect. Those three words are a decent technical spec for everything diffusion models are worst at:

• Imperfection — these systems are tuned toward the pleasing median. Ask for a flaw and you get a stylized, aesthetically pleasant flaw.
• Singularity — the business model is reproducibility at scale. One-of-a-kind is a bug in that economy, not a feature.
• Irreducibility — a prompt is a compression of intent into a few hundred characters. Compression means loss.

That’s a better review of generative image tools than most of what I read in tech press. Coming from the Bishop of Rome, that’s a little embarrassing for the rest of us.

Where I part ways with the Pontiff

I use these tools. They’re genuinely useful for storyboarding, for iterating on a visual idea before committing real hours, for people who have taste but no drawing hand. A strong critique of AI art as a replacement for human work isn’t the same as a critique of AI art as a tool inside human work, and the two keep getting collapsed into one argument.

The “spark of humanity” framing also runs into an awkward question: whose spark? Plenty of human-made commercial art is cynical, derivative, and produced on deadline by someone who stopped caring three clients ago. The spark is not evenly distributed among humans either. The honest version of this argument is about authorship and consent, not metaphysics.

Which brings us back to the scraped data. That’s the part of his critique with actual teeth, and it’s a legal and economic problem long before it’s a spiritual one.

The deepfake problem is the real story

Here’s where I think the aesthetic debate is a sideshow. A diocese in Illinois has published advice on identifying AI fakes of Pope Leo. Think about what that means operationally. A religious institution with over a billion adherents now has to run media verification as a pastoral function, because synthetic video of its leader spreads faster than the correction.

Bad AI art is an annoyance. Bad AI attribution is an attack surface. One of them clutters your feed. The other one lets anyone put a doctrinal statement, a political endorsement, or a provocation into the mouth of the most quoted religious figure on earth, and the only defense is that enough people learn to look closely at hands and audio artifacts.

The tool vendors shipping ever more capable image and video generation have built nothing meaningful to address this. Watermarking is inconsistent, provenance standards are optional, and detection is a losing arms race by design. The industry’s answer has effectively been “users should develop discernment,” which is a remarkable thing to say about a product you built.

My actual take

Leo XIV’s art critique is thoughtful and will change nothing. The deepfake problem is urgent and will change nothing either, until someone with legal standing forces it to. That’s the uncomfortable symmetry.

If you’re shipping generative media tools, the Pope just handed you a free product review with two findings. You’ll probably argue with the first one and ignore the second. The second is the one that will eventually show up in court.

🕒 Published: October 2, 2026