Painting with a Critic
I work mostly in acrylics and Acryla gouache, in a loose naive style. The paintings are my own. The process of getting them there now involves an AI. This essay describes how, and what has actually changed.
Why this matters for naive art specifically
Naive art is not trying to imitate academic realism or conform to formal technique. It is driven by instinct, narrative, and internal logic. That is a strength. It is also its failure mode. Without structured feedback, naive work can easily circle its own habits. It stagnates. The artist keeps producing the same piece, slightly differently.
What AI has given me is a way to break that loop without losing the rawness the style depends on. A mirror, a critic, a quiet collaborator. Not a substitute for the work, and not a generator of it.
That distinction matters, and the rest of this essay sits inside it.
What the AI doesn't do
It doesn't paint. It doesn't generate imagery that ends up on the canvas. It doesn't produce reference photos, underdrawings, or finished images I trace. It doesn't decide what I paint or why. None of my work is AI-generated in the sense that phrase usually means.
That needs saying clearly because the question gets asked.
What it does do
It acts as a critique partner. Not a flattering one.
I use Claude, an AI model made by Anthropic, as a sitting curator. I show it sketches, compositional studies, work in progress, and finished pieces. It responds in writing with structured critique: composition, colour, line, narrative, execution. What works, what doesn't, what to decide before going further.
The value is not in the AI having taste. It doesn't, not really. The value is in a consistent, attentive interlocutor that will read a sketch carefully, ask what the flowers are for, notice that a figure is floating without a ground plane, point out that a cloud above a setting sun reads accidentally as a mushroom cloud, and tell me honestly when a wing overlaps competing planes. I can get those observations from a human critic, but not every day and not at three in the morning.
How a painting actually comes together
Take Avis Aeromechanica Chocolatus, a 1000 x 750mm Acryla gouache piece I finished recently.
It started as a pencil drawing of a winged fuselage hybrid, a bird-aircraft with chocolate tanks and labelled parts. I had built the field plate first: a taxonomic document complete with Flight GPT, rat cargo, and a stowaway at the fuel inlet. That world-building is mine. The deadpan invention is where my voice lives and I wouldn't hand that over.
When I was ready to turn it into a painting, I showed the drawings to the AI. The exchanges went through several passes.
First pass. Concept critique. The flower foreground needed a reason to be there. The wing position needed committing to. The fuselage windows were pushing the image toward commercial aviation rather than mythological creature.
Second pass. After I resketched, it told me plainly that I had gained velocity but lost the world. The environment had vanished. That was useful. It stopped me transferring a half-resolved composition onto canvas.
Third pass. On the cliff-top viewpoint. Horizon placement, cliff edge as foreground anchor, where the viewer is actually standing.
Transfer. Grid method, horizon first, cliff edge second, fuselage, wings last. That order came from the critique, not from me.
Underpainting, blocking, staged build-up. I photographed and shared progress at the 50%, 70%, and 90% mark. Each time: what's resolved, what still needs work, the single strongest edit remaining. When I placed a small snail on the harbour wall as a deadpan taxonomic detail, it recognised the Insect Hotel logic and said nothing more needed adding. When the lighthouse base read as muddy, it said so.
At the final stage, it reviewed the piece against a rubric. Composition, colour, concept, execution, supporting elements, narrative atmosphere. Gave me a score and noted residual niggles I could live with. I did. The painting is listed now.
What I am actually learning
Three things have shifted.
Composition, learned functionally. I do not approach composition through the golden ratio or classical perspective systems. I work through adjustments. Shifting elements to create tension. Introducing asymmetry to increase movement. Adding or removing objects to control visual weight. When I asked whether to add a light beam from a lighthouse to a passing aircraft, that was not a technical question. It was compositional and narrative at the same time. The beam would connect two separate elements, introduce direction and flow, and make the lighthouse an active agent in the scene. That kind of reasoning accumulates. Each decision is tied to meaning, not aesthetics in the abstract.
Narrative, treated as the discipline. My paintings are not static images. They are unresolved situations. Something is always slightly off. A familiar environment behaving strangely, objects interacting with intention, elements that hint at a larger unseen system. The critique helps me make those narratives more legible without explaining them away. Too much clarity and the work becomes literal. Too little and it becomes incoherent. The work is in knowing which elements carry the story, which distract from it, and where ambiguity is productive rather than just noise. This is where naive art becomes sophisticated. Not through technique, but through control of meaning.
Iteration, treated as skill acquisition. The loop is simple. Create, present, critique, adjust, repeat. Because the AI responds immediately and consistently, I can compress what would be a long feedback cycle into a tight one. Multiple critique sessions in a single sitting. Over time that builds pattern recognition. I anticipate compositional issues before they are flagged. I recognise when a piece needs restraint rather than addition. I develop an internal sense of balance and tension. That is skill, even without formal training.
Who does what
I do the painting. All of it. The concept, the sketches, the mixing, the brushwork, the world-building, the decisions about what the work means.
The AI does not create. It reads what I show it and writes about what it sees, in the voice and register I have asked for. Curator-style, rubric-led, honest, no flattery. It asks clarifying questions when the intent is unclear. It notices things my eye has stopped noticing because I have been looking at them too long. It holds me to decisions I have already made and queries the ones I am about to drift from.
Preserving the naive
This is the outcome that matters most.
There is a common failure mode when self-taught artists seek improvement. They begin to replace their natural style with borrowed conventions. The work becomes more correct and less distinctive. The naive voice gets traded for something blander and more competent.
That has not happened here. My compositions are becoming more deliberate. My narratives are becoming clearer but still open. My use of space and elements is more controlled. But the core remains. Flattened perspectives. Symbolic rather than realistic proportions. Emotionally driven colour choices. A world that feels familiar but slightly unstable.
The work is still naive. It is just no longer accidental.
Why this works
The AI has no social stake in the outcome. It will not soften feedback to protect a relationship. It says the rocky foreground feels like a separate stylistic zone if it does. That honesty is hard to find in any studio conversation.
It is available at the speed of the work. Paintings move in hours and days. A critic in the room on demand is a different tool from a critic you see occasionally.
It scales well to technical choices I do not want to relitigate every time. Gesso ground for acrylic at this scale. Underpainting tone to carry pink and grey feathers. Grid method for banking angles. Those conversations are settled now. They are one prompt away.
Honest limitations
It does not replace a human eye in front of the painting in natural light. It cannot tell me how the surface feels, whether the varnish is uneven, whether the scale changes when you stand three feet back. It works from the photographs I send. It is assessing a digital reproduction, not the object.
It can also be too agreeable, occasionally. I have learned to ignore early drafts that read as helpful rather than exact, and to ask again with firmer framing. "Critique" gets better results than "what do you think".
A note on authorship
The work is mine. The AI is a tool in the same category as a good critic, a studio visit, or a book on composition. It shifts what I notice and when. It does not decide what the work is.
What I have built with it is not skill replacement. It is structured self-awareness. There is no transfer of creative ownership. Only an increase in clarity and intention.
If that distinction matters to you as a viewer or a buyer, I think it should. It matters to me.
This is a personal essay. The views are my own.