Zero. That is how many pixels in a diffusion model’s output were ever touched by a hand that held a brush. Every one of them is a sample drawn from a probability distribution learned from scraped data. I say that as someone who builds these pipelines for a living, not as a critic throwing stones from outside.
Which is why the recent critique from Pope Leo XIV landed differently for me than most celebrity takes on generative AI. He argued that AI-generated art lacks the “spark of humanity” that authentic human creativity possesses, and that algorithms built on probabilistic patterns derived from scraped data fundamentally lack the essence of art. Strip the theology out of that sentence and you have a reasonably accurate description of how latent diffusion works.
A technical claim in vestments
Most high-profile objections to AI art get the mechanism wrong. They describe collage, or theft, or a magic box that “knows” things. The Pope’s framing does not. Probabilistic patterns over scraped data is, more or less, the architecture. You train on a corpus, you learn a distribution, you sample from it with conditioning. The output is plausible because plausibility is the objective function.
That is the part worth sitting with if you write this code. Our models optimize for likelihood. Nothing in the loss function rewards a decision that a human made on purpose, for a reason, at cost. There is no term for intent. We have gotten very good at producing images that look like the kind of image a person would make, which is a different target than making the thing a person wanted to say.
You can disagree with the conclusion and still concede the mechanism. I do, partly. I think a human directing a model with taste and iteration is doing something real. But “I prompted it” and “I made it” are not the same sentence, and pretending otherwise has made the whole conversation dumber.
Imperfection as a product requirement
Commentary on Leo XIV’s position, including a February 2026 analysis of his remarks, frames his concern as wanting the arts to safeguard imperfection, singularity, and the irreducibility of human experience. Read that as a spec and it is a strange one, because those three properties are exactly what our systems are built to sand off.
Think about the defaults in any generation stack:
- We tune guidance scales until outputs stop being weird.
- We add negative prompts to suppress anything unusual.
- We run upscalers and face-restoration passes that pull features toward the statistical average.
- We generate four variants and pick the cleanest, which usually means the least specific.
Every one of those is a vote against singularity. Imperfection in a generated image is noise to be filtered, not signal to be kept. If you have ever watched a model turn a genuinely odd composition into a glossy stock photo after one refinement pass, you know the feeling. We built that behavior deliberately, because clients asked for it.
The deepfake part is squarely our problem
There is a second thread here that builders cannot wave off as philosophy. AI-generated deepfakes have circulated misrepresenting Pope Leo XIV’s words, to the point that the Diocese of Springfield in Illinois published guidance on spotting fakes of him. A local diocese is now running media-literacy education because our tooling made impersonation cheap.
When a religious leader with a billion-plus adherents becomes a reliable target for synthetic quotes, the failure is not a content-moderation failure downstream. It is an architecture failure upstream. We shipped generation at scale and treated provenance as someone else’s roadmap item.
Practical things that belong in your stack today, not in a future ethics review:
- Attach signed content credentials at generation time, not as a post-hoc watermark that survives one screenshot.
- Log the prompt, model version, and seed for anything you publish, and keep that log retrievable.
- Label synthetic media in the UI where users actually look, which is the feed, not a settings page.
- Refuse generation of identifiable public figures making statements, and treat that refusal as a feature your users can see.
- Build verification tooling, not just generation tooling. Detection is less fun to demo and more useful to ship.
What I am taking from it
I am not giving up image models. I use them in bot projects weekly, and they solve real problems for people who could never afford an illustrator. But I have stopped describing output as art in my own docs, because the word was doing work it had not earned, and because a fairly precise critique of my tooling arrived from an unexpected direction.
The Pope is describing a limitation we encoded on purpose. We optimized for the average and called it quality. That is a design decision, and design decisions can be revisited by the people who made them. That is us.
🕒 Published: