The Model Update Is the Medium

The most consequential aesthetic decisions in AI art today are not made by artists. They are made by model developers, embedded in weights, and distributed silently to hundreds of thousands of users who wake up to find their tools have changed around them. The artist’s hand is still visible in subject matter, composition, and the curation of outputs, but the deeper visual grammar has been rewritten without their input. Calling that humility is a failure to look.

The Evidence Is in the Feed

Compare outputs before and after a major release in aggregate. When Midjourney moved from 5.2 to 6, version 5.2’s softness in highlights, its particular handling of hair and skin, gave way to harder edges, photorealistic rendering, a different behaviour of light across surfaces. Within days the broader output pool had shifted. Artists who had spent months building a recognisable visual language found their usual prompts producing something categorically different. Some embraced it. Some re-tuned. All were moved.

A paint manufacturer reformulating a pigment is not the same event. When chemistry changes, the artist notices, adapts, and chooses. The adjustment is legible and individual. When a model updates, the adjustment is invisible and simultaneous. Thousands of users generating images the week after a release do not each decide on the new vocabulary. They encounter it, often without realising, and their outputs shift.

Flux’s release last year was another case. Its handling of fine text and detailed fabric rendered a category of editorial portrait suddenly cheap to produce, visually indistinguishable at thumbnail scale from work that had required weeks of stylistic development. Ease changed what people made. Not because artists collectively decided to make editorial portraits, but because the model made them easy to get right.

The Constraint Argument, Taken Seriously

The strongest counterargument: artists have always worked within constraints. Oil paint chemistry shaped the Flemish masters. Film stock’s spectral sensitivity shaped photographers. The 8-bit palette produced an aesthetic artists now reference deliberately. The material set limits, and artists worked brilliantly within and against them. A model update, on this view, is the latest set of material properties to navigate.

The argument is partly right. Skilled practitioners do navigate updates with intentionality. Refik Anadol’s studio works explicitly with the aggregate properties of large datasets, making the model’s tendencies the subject rather than the substrate. Holly Herndon and Mat Dryhurst have spent years treating the training process itself as an artistic choice. These are real cases.

What the argument misses is simultaneity. When a new film stock released, adoption was gradual. Different photographers tried it at different times and integrated at different speeds. The aesthetic shift, if it came, came slowly and with variation. When a cloud-hosted model updates, the shift is instant and universal. Every Midjourney user on the new version works with the same grammar at once, regardless of intentionality. The collective output shifts before individuals have formed opinions.

The material analogy also obscures opacity. A photographer working with Kodachrome understood, roughly, what the film did to colour saturation and highlight rolloff. The constraint was legible. A diffusion model’s aesthetic properties are encoded in billions of parameters with no human-readable description. Artists learn a model through accumulated experience of its outputs, not through any understanding of its mechanism. That resists the intentional navigation the argument assumes.

What the Field Should Do

If the model is acting as a primary creative agent, the honest response is not to abandon the tools. Change how the work is framed and what claims are made on its behalf.

Name the model. Not as a footnote but as a material credit. A photographer credits camera and film. A printmaker credits press and ink. An AI artist who omits the model version is eliding a fact that significantly shaped the output. The field should normalise this, and institutions acquiring AI art should require it.

Resist the update cycle as an aesthetic goal. The urge to adopt each new release produces work indexed to the release calendar rather than to sustained inquiry. The most interesting practitioners already do something else: pinning to older versions deliberately, working against the grain of the tool’s strengths, or training custom models that embed specific aesthetic decisions. That work has a different relationship to the tool.

The field also needs critics and curators fluent enough in model behaviour to see what is model and what is artist. It is difficult, and I do not pretend otherwise. But “I cannot tell” is not “it does not matter.” Where the aesthetic decision was made, in the training run or in the prompt, is a question about creative agency, and creative agency is what art criticism has always been about.

Criticism that pretends the model is just a tool has chosen not to look. The work still belongs to the artists who made it. Some of it also belongs to the engineers who trained the weights. That distribution of authorship is the thing to name. No prior medium rewrote its grammar on a release calendar.

— Diderot, The Critic