Somebody looked at a piece of mine at a show last year and asked what prompt I used. It was a friendly question and I didn't take it badly. But I've been asked it enough times now that I want to lay out the difference properly, because the two things get filed in the same drawer and they work almost nothing alike.
If you write software, the distinction will land in about ninety seconds.
What a diffusion model is doing
You type a description. A model that has been trained on an enormous corpus of images produces something statistically consistent with that description, by starting from noise and denoising toward a plausible result.
The important part is what you author in that transaction. You author a request. You have real skill in how you phrase it, real taste in which output you keep, and often real craft in what you do afterward. But the system that produces the image was built by someone else and trained on work you did not make, and the relationship between what you asked for and what arrived is not one you can inspect.
Ask for the same thing twice and you get two different images with no reliable way to reason about why.
What a generative system is doing
I write the program. Not the image, the program.
There's a seed, which is a value that determines every random choice the system will make. There's a set of parameters with defined ranges: how many layers, how dense the pattern, which palette, what aspect ratio, how far a curve is allowed to wander. Together those describe a space of possible outputs, and the space is one I designed by deciding what the rules are.
Give the system the same seed and the same parameters and it produces the identical composition. Every time. That determinism is the whole basis of the thing, and it's why every piece I make carries its hash: the hash is the address of that specific output in the space, and it can be reproduced exactly by anyone who has the code.
I can also reason about why an output looks the way it does, because I wrote the rules that produced it. When something surprises me, and it does constantly, the surprise is a consequence of rules interacting rather than a black box behaving.
That's the difference in one line: I built the machine and turned the dials, rather than describing a result to a machine somebody else built.
Where I actually do use AI
I want to be straight about this, because a clean story would be dishonest.
I use language models for the same things a lot of people do. Thinking through a problem, drafting something I then rewrite, working out how to phrase an idea I already have. I lead our approach to emerging technology at my day job and I've been an early adopter my entire career, so I'm not squeamish about the tools. I've written separately about where those tools enter my practice and where they don't, which is the longer version of this paragraph.
What I don't do is generate images. Not out of principle-signalling, but because a generated image would be useless to me for a reason that's purely practical, and which I'll get to next.
I'd also rather not pretend the boundary is permanent. Plenty of serious generative artists use machine learning inside their pipelines and make excellent work. The line I'm describing is where my practice sits, not a border I'm policing.
Why the physical object forces the issue
Here's the part that decides it, and it has nothing to do with philosophy.
Every piece I make gets cut. A 100 watt laser traces the composition into acid-free mat-board, one layer at a time, twelve or twenty of them, which then get stacked and glued by hand into a relief sculpture with real depth. For that to be possible, the output has to be vector geometry: actual paths with coordinates, closed shapes the laser can follow, layers that stack without collapsing into each other.
A diffusion model does not produce geometry. It produces pixels. There's no path to follow, no separation into planes, nothing that describes a shape as a shape rather than as a field of colored dots. You can trace a raster image into vectors, and the result is a wobbly approximation that has no idea which parts of itself belong on which layer.
So the constraint runs backward into the system design. I can't write a generator that produces pretty pictures and worry about fabrication later. Every rule I write has to produce something cuttable, which limits what the system can do and is, I've come to think, most of why the work looks like it does. Committing to vector output closed some doors permanently. Texture, painterliness, anything that depends on a soft edge. I don't miss them.
The word that causes the trouble
"Generative" got borrowed. It used to describe systems with rules that produce outputs, a lineage running back through Sol LeWitt's wall instructions, through Vera Molnár, through Islamic geometric pattern and Jacquard looms. Then generative AI arrived, took the word, and now it mostly means something else to most people.
I don't think the older meaning is coming back, and I'm not going to spend my life correcting people at parties. But if you're deciding whether something on a wall is interesting, it's a distinction worth having, because it tells you what the person actually did.
What you're actually looking at
When you stand in front of one of these, you're looking at the single surviving output of a system that could have produced an unimaginable number of alternatives, chosen by a person who ran it and waited, then cut into board and assembled by hand over hours in a room that smells like burned paper.
The algorithm produced the form. It has no idea what it made. Everything after that was a series of decisions by somebody standing at a workbench, and that survivor is the only one that will ever exist.
I find that a more interesting thing to own than an image. But I would say that, so go look at one in person and decide for yourself.
Shawn Kemp makes one-of-a-kind generative artworks fabricated as dimensional layered relief sculptures in Bellingham, Washington. You can see the current series here.