Beautiful But Wrong

AI-generated architectural visuals are becoming more impressive by the day. Sometimes it's nearly impossible to tell them apart from real photographs at first glance. But here's what I've come to realize: no matter how beautiful a visual is, if it's technically wrong, it's completely useless.

Staircases that lead nowhere. Balconies without railings. Spaces that are ergonomically impossible. Objects with incorrect scale. I see these every day. AI can produce an aesthetically satisfying image, but it has no way of knowing whether what it's created could actually be built in the real world. Only someone who truly knows the field can catch that.

There's an important distinction here: the intended use changes everything.

For visuals used during the concept or decision-making phase, these kinds of errors are tolerable. What matters there is communicating an idea, not technical accuracy. But when it comes to sales visuals for a real project, the situation is very different. The client looks at that image and says "yes, this is what I want" — and makes an investment decision based on it.

At that point, a beautiful but technically wrong visual can turn into a serious financial or contractual dispute between the developer and the client.

In my own workflow, AI plays a role at both the process and output stages. I use AI to test framing and lighting alternatives first — comparing morning, midday and evening light to see which one best captures the feeling I want to convey. When a project needs a night version, I use AI to explore ideas on top of a daytime render. In a sense, I use AI as a draft tool — a way to test an idea before committing to it. But ultimately, I build every project's 3D model myself, take the render myself, and verify the technical accuracy myself. AI adds speed and aesthetic refinement; I build the foundation.

For anyone using AI as a tool: without a technically trained eye to verify what you're producing, that tool can sometimes leave you with nothing more than a beautiful mistake.