The popular anxiety about AI art has always been framed wrong. Critics worry that anyone can now make something that looks good, which devalues the work of trained artists. That's true, and it's also beside the point. What's actually happening is a bifurcation: a mass of single-prompt tourists on one side, and a much smaller group of practitioners who've figured out that the interesting work starts where the prompt ends.
Refik Anadol is the clearest case. His 2024-2025 'Large Nature Model' installations, shown at MoMA, the Serpentine, and the UN, weren't made by typing something clever into a text box. He directs a 15-person team across a multi-stage pipeline: curating datasets from the Smithsonian and the Natural History Museum, training a custom diffusion model, then iteratively tuning parameters and aesthetics until the output matches a vision. He's said he is 'involved in setting every single parameter.' That's not prompting. That's directing, in the full filmmaking sense of the word.
The Real Skill Is the Workflow
Tools like ComfyUI have made this kind of multi-stage control accessible without a 15-person team. Its node-graph model lets artists chain pipelines across multiple models, stack LoRA fine-tunes, apply ControlNet for pose and depth guidance, and run multi-stage upscaling, all as a repeatable, adjustable production workflow. Industry practitioners are pretty blunt about what this means: it's what separates amateur AI art from professional production work. The learning curve is steep enough that it functions as a genuine skill barrier, which is exactly what the 'anyone can do this now' crowd tends to miss.
Formal research is catching up to what practitioners already know. A peer-reviewed paper presented at CHI 2025 studied LACE, a tool built specifically for professional visual artists that supports iterative, human-in-the-loop refinement across multiple steps. The researchers drew a pointed contrast between clunky sequential workflows and more sophisticated parallel refinement. The fact that HCI researchers are now publishing on multi-step iteration as the distinguishing skill for serious AI artists is a signal that this has moved past craft folklore into something people are treating as a structured discipline.
Adobe Firefly's numbers tell the same story from a different angle. The platform crossed 24 billion cumulative generated assets by June 2025, with traffic growing more than 30% quarter over quarter. The figure that actually matters, though, is session length: 26 minutes on average. That's not someone generating an image, shrugging, and closing the tab. That's someone working, adjusting, comparing, and iterating. It's a usage pattern that looks a lot more like Photoshop than like a novelty toy.
This Doesn't Rescue Every Argument About AI Art
To be fair to the skeptics: none of this resolves the legitimate arguments about training data, consent, and the economic displacement of illustrators and concept artists. Those problems are real and the industry has handled them badly. A sophisticated workflow doesn't launder an ethically murky foundation.
But the specific criticism that AI art requires no skill, that it's just automated output anyone can produce, is getting harder to sustain. The ceiling keeps rising. The artists operating near it are doing something that looks much more like curation, direction, and iterative craft than it does like typing. Prompting was never the point. It was just the door.
