There’s a question that has been circling generative art since image models became good, and it got a public airing this month: if the output is a picture, and nobody can tell from the picture how it was made, does hand-coding still mean anything?
Piter Pasma wrote about it for Right Click Save in September 2026, prompted by an Art Blocks discussion on Instagram on September 2. Pasma is a reasonable person to hear from on this. He’s a Dutch generative artist whose reputation rests on mathematical density rather than visual style — Skulptuur for Art Blocks Curated, and Universal Rayhatcher, which renders raytraced scenes from extremely compact code. His work is the kind that stops being impressive if you learn it was prompted.
Why the question is harder than the usual answer
The comfortable reply is “process matters, provenance matters, the artist’s intent matters” — all true, all unfalsifiable from the artwork alone, and none of it survives contact with a viewer who simply doesn’t care how a thing was made.
The more useful framing is about what the two methods can actually promise.
A hand-coded generative system is deterministic and enumerable. Given a seed, it produces exactly one output, every time, forever. The artist can reason about the entire output space — not sample it, reason about it — because they wrote the rules that define it. A long-form collection of 1,000 works is a claim that the artist understood the space well enough to let a hash function pick 1,000 points in it and stand behind all of them.
A diffusion model offers none of that. It’s a sampler over a distribution nobody fully characterised, including the people who trained it. You can curate its outputs beautifully. You cannot make the long-form promise, because you cannot inspect the space.
That’s a real, technical difference, and it’s the one that doesn’t dissolve under “but the picture looks the same.”
Where it gets uncomfortable
Two places, and it’s worth naming both rather than pretending the distinction is clean.
First, the boundary is already porous. Plenty of working generative artists now use coding assistants to write the code that generates the work. Is a hand-coded piece still hand-coded if a model wrote the shader? Most people’s instinct is yes — the artist specified, evaluated and iterated — but that instinct is doing the same work as “the artist prompted, evaluated and iterated,” and it’s not obvious where the line falls.
Second, the argument proves less than its defenders want. Determinism and an enumerable output space are real properties, and they are technical properties. They explain why hand-coded generative art can make promises AI art can’t. They do not, by themselves, make it better art. A boring deterministic system is still boring.
What this is actually about
The reason this keeps coming up in the generative art world specifically — rather than in painting, where the same anxiety exists — is that generative art’s entire legitimising story was the algorithm is the artwork. Vera Molnár, Frieder Nake, Manfred Mohr: the system is the thing, the print is a readout.
If the readout can now be produced without the system, the field has to say out loud what it always assumed was obvious. Pasma’s piece is part of that, and the fact that it’s being written by someone whose work is unusually algorithm-dependent is not a coincidence.
The version of this argument worth making isn’t defensive. It’s that a system you can reason about is a different kind of object than a system you can only sample, and that difference is worth preserving on its own terms — not because it fends off AI, but because it was always the interesting part.
Related Reading
- How can generative artists distinguish hand-coded art in the age of AI? — Right Click Save
- Space #51 — Endless Possibilities with Piter Pasma (Universal Rayhatcher) (YouTube)
- The Interview | Piter Pasma — Right Click Save
- fx(hash) in Venice — Piter Pasma’s ‘Industrial Devolution’ (YouTube)
- Universal Rayhatcher — Art Blocks