You are standing in a line somewhere, phone out, thumb moving. An image of Pope Leo XIV appears in your feed. The lighting is beautiful. The vestments drape correctly. There is a caption attributing words to him that sound almost right. You pause for maybe half a second, decide it’s probably fine, and keep scrolling. Multiply that half-second by a few hundred million and you have the actual state of information in 2026.
That scenario is why the Pope’s recent criticism of AI-generated art caught my attention, and not for the reason most coverage suggests. He has argued that machine-produced visuals lack the “spark of humanity” found in authentic human creativity, and that algorithms depending on probabilistic patterns pulled from scraped data miss something essential about art itself. Commentary around his position frames the arts as a place to protect imperfection, singularity, and the irreducibility of human experience. Meanwhile, AI-generated deepfakes have been putting words in his mouth, which gives the whole argument an unusually direct demonstration.
I read theological statements about technology the way I read a well-written bug report from a non-engineer. The vocabulary is not mine. The observation is often correct anyway.
Probabilistic patterns are not an insult, they are a spec
When the Pope describes generative models as working from probabilistic patterns over scraped data, that is not a slur. That is roughly the architecture. A diffusion model learns a conditional distribution over images and then samples from it. The training objective rewards outputs that look likely given everything the model absorbed. Likelihood is the whole point.
Now hold that next to the claim that art should protect imperfection and singularity. Those two things are in mathematical opposition. Singularity means low probability under the training distribution. A one-of-a-kind brushstroke, an oddly chosen color, a composition that should not work but does, these are outliers. A system tuned to produce plausible output treats outliers as error. You can crank up temperature or classifier-free guidance and get weirdness, but that is sampled weirdness, drawn from the same learned manifold. It is not the weirdness of a specific person who lived a specific life and made a specific irreversible decision at 2 a.m.
So when someone says there is no spark, I would translate it as this: the generative objective has no term for the thing you care about. Not suppressed. Not approximated badly. Absent from the loss function.
Why this matters for agent design and not just for art critics
Readers here mostly care about agents, so let me connect it. The same property that makes generative art feel hollow makes agentic systems dangerous in a specific, predictable way.
An agent that generates content is optimizing for plausibility. It is not optimizing for provenance, attribution, or truth. Those are separate concerns that have to be built as separate components, and in most deployed stacks today they simply are not there. A deepfake of Pope Leo XIV saying something he never said is not a failure of the image model. The image model succeeded. It produced a highly probable image given the prompt. The failure lives one layer up, in the absence of any architecture that tracks where a thing came from.
This is the part I think gets lost. The response to synthetic misrepresentation cannot be better taste or more careful viewers. Human discernment does not scale to feed velocity, and the half-second pause I described above is the realistic budget. What has to scale is structure:
- Cryptographic signing at the point of capture, so authentic material carries a verifiable claim rather than asking audiences to detect fakes
- Provenance chains that survive re-encoding, cropping, and screenshot laundering, which is where most current schemes quietly break
- Agent pipelines that treat generated artifacts as untrusted by default and refuse to attribute statements without a source they can check
- Institutional publishing channels that audiences can verify directly, so the authoritative version is cheap to find
Note that none of these require resolving the philosophical question. You can believe generative models produce real art and still want a signing layer. The engineering follows from the architecture, not from the aesthetics.
What I take from a critic who does not write code
The tension being described, traditional craft on one side and algorithmic output on the other, is usually framed as a fight about jobs or taste. I think it is more interesting as a question about what our objectives actually encode. We built systems that are very good at producing the expected value of human expression. Then we are surprised when the output feels like an average.
Someone noticing that the spark is missing is giving us useful information about our own design choices. Whether we build systems that can hold singularity, or just systems that can prove who made what, both are real problems. We have barely started on either.
🕒 Published: