A painter I know — oils, landscapes, a small gallery in Bath — has been running a diffusion model through her process for about six months. Not to make her paintings. To decide which ones are worth three weeks of her life.
She uses it the way an earlier generation used Polaroids: a visual scratchpad. "It lets me see an idea fast, so I can work out whether I actually want to make it," she said. "Most of the time I don't. But when I do, I'm all in." The paintings stay entirely hers. The model just helps her say no faster.
That conversation has lodged in my head because it sits so far from the two stories everyone else tells about generative AI — the breathless one ("now anyone's an artist!") and the apocalyptic one ("it'll replace every artist!"). The reality, for the practitioners I'd actually trust, is less dramatic and far more interesting.
Where it earns its place
It's very good at exploration — the first twenty riffs on a design, the mood board, the thumbnail before the real work begins. It excels at quickly showing you adjacent possibilities you might not have wandered to.
It's good at fiddly finishing — a realistic texture, a filled-in background, the tourist removed from your photo. An hour of Photoshop collapsed into a second's prompt.
It's surprisingly good at breaking a block. Sometimes you just need something on the page to argue with. A bad generated draft earns its keep not because you'll use it but because it lets you name what you didn't want — which is half the job done.
And it does personalisation at scale that used to be impossible. For some products that's a real gift. For others it's the top of a slope I wouldn't want to be standing on.
Where it falls flat
It can't do intentional art. Models are averaging machines; they're magnificent at "make me something that looks like this genre" and genuinely bad at "make me something nobody's seen, with a reason behind every choice." The second sentence is what artists are for.
It doesn't know what to leave out, and the best work is defined as much by its absences as its presences. Models put everything in.
And it doesn't care — who sees the work, how it's framed, what it'll mean to one specific person in a room. That sounds sentimental. It's actually practical. You have to supply the caring, because nothing in the system will.
The trouble nobody in the industry likes to dwell on
Most large models were trained on text, images and music scraped from the open web — including the work of living artists who were never asked and are never paid. That's ethically murky at best and possibly illegal; several cases are grinding through the courts as I write. The artists I know are furious about it, including the ones who use these tools themselves. Both things can be true.
Then there's the bottom rung. Commissioned illustration, concept art, stock photography — that's how a great many creative careers start. If generative AI eats the entry-level work, we dismantle the training ground for the next generation of artists without anyone deciding to. No single studio can fix that alone; it needs an industry answer.
There's the pollution problem, too. A web where half the images and articles and videos are machine-made is a different web from the one I grew up trusting. We don't yet know what that does to discourse, to trust, to cultural memory. "It's probably fine" is not a finding.
And the misuse is real and getting worse — deepfakes, non-consensual imagery, automated propaganda. Watermarking and content provenance stopped being optional a while ago.
What the careful ones have in common
The creatives I watch using this well treat it as one tool among many, not a replacement for a practice. They credit it, in the work and in conversation, rather than pretending the tool wasn't in the room. They reach for it in the thinking phase — sketches, mood, exploration — and keep it out of the shipping phase. And where they can, they pay for tools with clean training data: Adobe Firefly trains on licensed material, and the open-source side is catching up.
We do use generative AI in our own work — early wireframes, code scaffolding, test data, the occasional first draft a human then takes apart. We try to be straight about where it shows up, and to keep the decisions that matter in human hands.
If you're working out where generative AI belongs in your own practice, come and think it through with us — and bring every ounce of scepticism you've got. The good conversations always do.
