Can I show you something you already know? Open a generative AI tool. Drop in a few reference images, type a prompt, hit go. Wait.
The output appears.
For a fraction of a second, there’s joy. A rainbow of colors. Crisp lighting. A real scene that starts to pull you into an idea you only half had. And then the logical brain kicks in. This dreamscape is not quite right. It’s missing the grit. The fine details that make it believable.
About 800 milliseconds into looking at it, you’re standing in a very familiar uncanny valley. We’ve all spent a lot of time down there lately. The valley may have finally banished the eleventh finger. It may now be able to show you a table without extra legs. But the polish never quite matches the detail, and creative eyes sit and gaze with irreverence at exactly that gap, willing it to read their minds without committing to a single minute of work.
Everything that happens after those 800 milliseconds is the real work. Not in the prompt. After it.
We Used to Sweat the Glimmer on a Button
It’s worth remembering how deep the web used to feel, because the comparison explains something.
Flash had its day, and interactive experiences were genuinely fun. Deep logic coding suddenly had room to breathe on a desktop. It had a canvas to spread out on. We made little buttons and UI that imitated tactile experience, and we spent real time on the perfect glimmer on a button. That effort was controllable, because it required raw human effort to exist at all. Nobody was going off and making things without you.
Then the web aged into all that tactility and we got a bit overwhelmed by the depth. We wanted simplicity. Microsoft handed us flatness with Metro, around Windows Phone 7 and Windows 8, and it reset design thinking toward something easier to take in. We moved faster. We had less astonishment about what we were doing. That was a correction more than a crime, but something did leave the room.
And now Apple has walked us back toward tactility with glass. It’s coming back, but it’s mathematic. Relentlessly predictable. It moves over time and space and it feels alive. That’s the wonder worth aiming at.
So the depth of craft didn’t disappear. It moved. And with AI it moved again, into a stranger job. You’re less a maker fine-tuning a thing and more a worried parent, reeling in the wrong directions, asking the only question that matters: what am I missing that’s a real tell this is fake?
Directing the Toddler
Here’s the problem we were handed. A property that doesn’t exist yet.
Pinnacle Village is a long build. All anyone had were architects’ drawings, and the question we needed to answer for a buyer was emotional, not technical: what would life actually feel like there? Leaving work. Going home. Getting ready. Meeting a friend at a restaurant downstairs.
Most realtors don’t have the budget for a full 3D animated exposition of that. Architecture assets don’t go that far. So AI was the solution, using Kling 2.0 for image-to-image animation and scene creation. Hours of image generation off the renderings. Lighting experiments. Not as many hours as a full 3D build, but honestly, with detail refinement and pulling faces back out of the valley, it could have been.
The method we landed on was to build a before and after image for every clip. Multiple attempts through Kling would net different, more refined results, but the thing that actually worked was giving it a real scene to start from and a matching real scene to end on, so there was less room to deviate. Thinking that way is like being a film director with Photoshop. Bring real elements into the scene. Make it believable. Then generate new angles to create the motion.
And the rules rewrote themselves for every single clip.
Which is the part people miss. The tool gave us no repeatable method. What carried the work was knowing in advance what AI would grab onto and obsess over:
- A human face in frame is always a risk, so we kept faces sparing and moving.
- Slow motion obscures whatever mistakes it wants to make, so we’d ask for an ultra slow-motion shot at eight seconds and speed it up or clip it in post.
- Cut from small, one-second acceptable pieces and assemble something believable, because AI is nowhere near composing an entire thing that works.
- Blurry objects are a trap. Anything out of aperture focus becomes an invitation to invent. Even Nano Banana still bullshits the objects on a shelf.
- Clean surfaces that behave predictably are key, which is a problem, because our own lived-in spaces are nothing like that.
Make the geometry easy to understand. That’s the job. It’s like directing a toddler on what reality is like. Every skill doing the actual work there — framing, continuity, lighting, knowing how a cut reads — existed long before the tool did. Someone without that background ships the bad version and never notices.
The Core Is Real, The Edges Are AI
Rogers-O’Brien has a thing they call R-O City. It isn’t a real place. It’s a 3D environment holding every project they’ve built over the years, large and small, and at combined scale it genuinely reads as a city.
We wanted the site’s footer to nod at it. So we built a fake-3D JavaScript parallax scroll, so that as you move you drop below the horizon of R-O City, or rise back above it on the way up, with a crane dressed in the R-O logo working away at the skyline. No OpenGL. No 3D engine. Photoshop first. We pulled assets out of their environment, reimagined the lighting, repositioned things by hand.
And around the edges, the parts almost nobody looks at, there is AI patchwork.
That’s how this should be used. The core of it, the focus, is real. The rest is filled and flattened so the spotlight lands where it should. It works because the eye is never invited to interrogate the thing that couldn’t survive interrogation.
Spitfire Bourbon runs on the same rule from the other direction. We were after WW2 Britain. Crusty tents littered with war correspondence and codebreaking, tough times taken off the edge with a bottle of spirits. Being a fighter pilot in those early days is arguably one of the most stressful jobs anyone has ever had. Even with a cutting-edge Merlin engine out in front of you, you had no idea which flight was the last one. We wanted the cross-section of the moment a nineteen-year-old, rattled from his first encounter with a Messerschmitt, takes a sip. Not an endorsement of the habit. But hell, wouldn’t you want a stiff drink?
That danger and adventure needed to come from somewhere real, so we generated from old open-source photographs, a real basis turned into something new for the brand. Real stories with our own spin on them. But the typography and everything else wasn’t left to AI. We fine-crafted that the old way. The good way.
Because that’s the focus. The focus should always be human.
Where the Line Actually Is
Plenty of serious designers are refusing AI outright on ethical grounds. Training data, scraped style, the whole argument. I understand it, and I’m not here to tell anyone their anger is misplaced.
But I think the ethics get decided by placement, not by abstinence.
If you let AI craft the focus, you’ve already failed. Human faces that need to carry actual emotion. A product that has to dazzle somebody. An entire composition handed over wholesale. That’s the problem, and it’s a judgment failure before it’s a moral one. If you use it to fill the gaps in an image, to make the crumbs on a countertop, the grain in a floorboard, that’s a different thing. How much IP is really wrapped up in the clouds on a particular afternoon in someone’s vacation photo? Did they own the clouds that day? Not likely.
Tarantino, who is more honest about this than most people in any creative field, says he steals from every movie ever made, and that great artists steal rather than do homages. He also described his own method as using the oldest chestnuts in the book and purposely having them run awry. That’s not a loophole. That’s the work.
And here’s the harder part, the part I’ll admit is easier to say while holding the prompt: artists were stealing from each other long before any of this existed. As a rule, everything always moves faster. Now it’s faster. The problem isn’t AI, it’s that some people don’t want to move fast, and that’s genuinely okay. But anyone who does want to participate in a world of fast-moving artists has to take a hard look at their own work and count how much of it came from things they saw, loved, and took.
Style Is a Liability
This is where I lose some of the room.
If you have a style, it can be stolen. That’s the liability, and there’s nobody out there protecting you. So you have two options. Become overly concerned with defending your IP, or become agnostic to that kind of protectionism by admitting the only real defense is to keep creating something new.
You can’t stop and collect a payout on the one good idea you had from now until retirement. That isn’t an artist. That’s a landlord.
Do we have a style we prefer? Of course. Everyone does. But we’re not beholden to it, and sometimes we want to breathe new air, stretch our legs, get challenged by something we haven’t made before. That’s normal. It’s also, right now, the only durable position, because a recognizable look is the single easiest thing in the world to reproduce. The chameleon survives. The signature gets scraped.
What This Means If You’re Buying It
Most of the terrible AI creative in the market comes from people who didn’t want to put effort into the work. They believed a few words in a prompt would give them something good enough.
You’ve watched the pattern. Somebody starts using AI for social content to promote their book or their channel. Backlash arrives. Then comes the video with the heartfelt apology and the promise to never use AI again, which is a little ridiculous, because the backlash didn’t come from the tool. It came from not taking the time to really review what they made, find the problems in it, fix those problems, and bring something real in first.
It’s just like any tool. It won’t do the work for you. You have to learn what it’s good at and what it sucks at, and that learning is painful and unglamorous and mostly happens in failed outputs nobody will ever see.
So if you’re on the buying side, two things.
AI is not a discount. It’s a redistribution of the same effort. The hours move from execution to direction, review, correction, and knowing where not to point the thing. What you’re actually purchasing is that judgment.
And we don’t use AI because a client asks for it. We use it because it fits our process. We set the process and the budget, then we decide what the work needs. That’s not stubbornness, it’s the only way the output stays worth your money.
New tools will keep landing on the desk. They always have. The rule hasn’t changed once: you have to put the work in.
