My favourite conversation about generative AI this year was with a painter — oils, landscapes, represented by a small gallery in Bath — who had been quietly using a diffusion model in her process for six months.
Not to make her paintings. To think.
She described using the model the way an earlier generation of artists described using Polaroids or early digital cameras: as a kind of visual scratchpad. "It helps me see an idea quickly, so I can decide if I want to spend three weeks actually making it," she said. "Most of the time I don't. But when I do, I'm more committed."
The paintings are still entirely hers. The tool just helped her choose which ones to make.
That conversation has stuck with me because it's so different from the dominant narrative about generative AI — which is either breathless ("AI makes anyone an artist!") or catastrophic ("AI will replace all artists!"). The reality, for the thoughtful practitioners I know, is quieter and more interesting.
What generative AI is actually good at
Exploration. The first twenty variations on a design idea. The mood board for a concept. The thumbnail sketch before the real work. Generative AI is very, very good at quickly showing you adjacent possibilities.
Finishing fiddly details. Rendering a realistic texture. Filling in a background. Removing a photobomber. These used to be an hour's Photoshop work; they're now a second's prompt.
Breaking creative blocks. Sometimes you just need something on the page to react to. A bad generated draft is often a useful launchpad — not because you'll use it, but because it lets you articulate what you didn't want, which is half the work.
Personalisation at scale. Generating unique content for each user (within reason) is now cheap. For some products this is lovely. For others it's a slippery slope.
What it's not good at
Intentional, meaningful art. Generative models excel at averaging. They're superb at "give me something that looks like [existing genre]." They are genuinely bad at "give me something I've never seen before, with a reason behind every choice." The latter is what human artists do.
Knowing what to leave out. The best creative work is defined as much by what's excluded as what's included. Models tend to put everything in.
Caring. This sounds sentimental, but it's practical too. Artists care about who sees the work, how it's framed, what it means to someone. Models don't. You have to bring that.
The honest trouble
I want to name the things the AI industry often glosses over.
Training data is often scraped without consent. Most large generative models are trained on images, text, and music pulled from the public internet — including the work of living artists who were not asked and are not paid. This is morally complicated at minimum, and possibly legally fraught. Several court cases are in progress. The artists I know are rightly angry about this, even the ones who use generative tools themselves.
The economic effects on junior creatives. Commissioned illustration, concept art, stock photography — these are how a lot of creative careers start. If generative AI eats the bottom rung, we lose the training ground for the next generation. Individual studios can't fix this alone; it's going to need industry-wide responses.
Synthetic content pollution. A web where half the images, articles, and videos are machine-generated is a different web than the one we grew up with. We don't yet know what this does to trust, to discourse, to cultural memory. It is not obviously fine.
Misuse. Deepfakes. Non-consensual imagery. Automated propaganda. These problems predate generative AI but are getting worse with it. Good technical watermarking and content provenance are not optional any more.
How the artists I trust are using it
A few patterns I've noticed among the creatives who seem to be using generative AI thoughtfully:
- They treat it as one tool among many. Not a replacement for their practice. An addition to it.
- They credit it. Both in the work itself and in conversation. Nobody pretends the tool didn't exist.
- They use it in the thinking phase, not the shipping phase. Sketches, mood, exploration — not the final piece.
- They pay for tools with clean training data where they can. Adobe Firefly, for example, is trained on licensed data. Open-source alternatives are catching up.
A note on our own practice
We use generative AI in our work — for early wireframes, for code scaffolding, for generating test data, sometimes for first-draft writing that a human then edits. We try to be honest about it. We try to ensure humans make the meaningful decisions.
If you're working through how generative AI fits into your own creative or product practice, we'd love to help you think. Book a chat — and feel free to bring your scepticism.
