I've watched three different illustrators try to use Midjourney in the last six months. Same pattern every time. They open it confident. They leave frustrated. Not because the outputs are bad, exactly... but because they spent forty-five minutes describing a feeling they could have sketched in thirty seconds.
Stanford researchers put language to what those illustrators were experiencing: the friction isn't technical. The tools are forcing visual thinkers to encode their intent as text, which is not how visual intent actually works. That translation cost is enormous. And most people just blame themselves for not knowing the right prompts.
Look. Here are the things I keep coming back to.
-
Prompt engineering is a skill that has nothing to do with being a good artist.
This is the part nobody wants to say out loud. You can be a genuinely skilled image-maker and still be terrible at prompting. The two abilities don't overlap much. Knowing to type "cinematic lighting, 35mm, f/1.8 bokeh, muted earth tones" is a technical vocabulary about describing pictures... not making them. Asking a painter to master that is like asking a composer to get good at music XML before they're allowed to write a melody. It's backwards.
-
The tools that actually help artists are the ones that stay in the artist's native language.
Adobe Firefly inside Photoshop works better for working designers not because the model is superior... but because the input is already a canvas with context. You're pointing at a thing and saying "fix this" or "extend this." That's a conversation. Midjourney from a blank text box is a cold call to someone who doesn't know you. The difference in creative flow is not subtle.
-
"Iteration" means something different to a tool than it does to a person.
When a human collaborator iterates with you, they remember the direction. They understand that when you said "warmer" last time, you didn't mean just color temperature, you meant emotional register. AI tools iterate statistically. Every prompt is a new conversation with something that has no memory of what you were trying to build. Some tools are patching this with persistent style references and character consistency... but that's still not the same as a collaborator who gets it.
-
The artists using these tools well have mostly stopped using them for ideation.
This is the thing I keep noticing. The illustrators, photographers, and designers who seem genuinely comfortable with AI tools have quietly stopped using them to generate ideas. They use them for execution of very specific, already-decided things. Background fills. Texture passes. Reference variations once a concept is locked. The tool becomes a production assistant, not a creative partner. That's a significant step back from the pitch.
-
The interface problem is real and almost nobody is fixing it.
Runway is interesting here. So is Krea. Both are trying to give visual artists an input method that doesn't require typing a paragraph of descriptors before anything happens. Krea's real-time canvas... where you sketch and the model responds as you draw... is genuinely closer to how visual thinking works. It's not perfect. But it's the right direction. Most tools are still optimizing the model and ignoring the interface. That's where the real work is.
-
"Good AI assistance" for visual work probably looks a lot less like generation and a lot more like response.
Here's the thing... the human collaborator model that works isn't someone who creates things for you. It's someone who reacts to what you make. Who says "that composition is fighting itself in the lower left" or "the palette shift between panels three and four is doing something interesting, should we push it?" AI that can analyze and respond to existing work is more useful to most artists than AI that generates from scratch. We have the generation part. We're still waiting for the response part to catch up.
I don't think these tools are failing artists. I think they were built by people who solved a different problem than the one artists actually have. Generation at scale, for people with no visual training, is a real thing to build. But that's not what a working illustrator needs at 11pm trying to hit a deadline.
The gap between those two use cases is bigger than anyone admits. And it's probably where the next useful thing gets built.