My kid took a photo of me the other day.
Nothing planned. I was in a black suit, standing against the wall in our living room, and she framed it and pressed the button. Behind me you could see a desk with the day's work still on it, a chair, the floor tiles. An ordinary room on an ordinary evening.
I liked it. But I wanted to see what it could become, so I took it to an AI image model and asked for a professional studio.
The first thing I learned was what the machine does when you're not looking
The studio came back beautiful. Seamless backdrop, soft key light, a proper floor. Everything I asked for.
But the man standing in it wasn't quite me.
The eyes were slightly off. Not dramatically, not in a way most people would catch, but I caught it immediately, because it's my face. Then the lips. Then something around the cheeks. Each pass I ran, the model would fix one thing and quietly redraw another.
So I did what you do. I wrote a better prompt. I told it plainly: change nothing about the subject, not a single pixel, only the lighting and the colour. No guesswork.
It changed me anyway.
Here's the part worth understanding, and it took me a while to accept: this isn't the model disobeying. Generating an image means redrawing every pixel. There's no setting for "leave this face alone," because leaving something alone isn't an operation these systems perform. Asking for it is like asking a painter to photocopy.
And what it does when it thinks it knows you
There's a second thing I noticed, and I'll say it carefully, because it's easy to take it the wrong way.
The model didn't make me look like someone else. It made me look like a type. Features pushed further than they sit on my own face, pulled toward some average it had learned from whatever pictures it was trained on. Not a person in a photograph. A category wearing my clothes.
Every round, I found myself dialling its assumptions back down to arrive at the man who was actually in the room. Which is the same lesson I keep writing about here. The machine was generalising. It had a category, and it applied the category, because a category is cheaper than looking closely.
I know that habit well enough. Most of us do. Being filed under whichever label is convenient, then having to work your way back out of it. Generalising is the easy route, the cheap route, and it keeps failing us.
It turns out software learned the same bad habit from us. Where else would it have learned it?
So I turned the automation down
The fix wasn't a better prompt. It was less AI.
I let the model do what it's genuinely good at — building a room that never existed, with light I don't own the equipment to make. Then I stopped letting it touch my face, and put my real face back by hand: aligning it to the plate, keeping the studio's light but my own features, matching my skin to my own neck, adding one soft light that stops at the edge of my body, grading it, and stopping.
Every step where I reached for the automatic option, I got something smoother and less true. Every step where I took the manual control, I got something that looked like a man in a photograph.
There's a whole craft in knowing which is which, and it's the same craft in every field: knowing what to delegate and what to keep your hands on.
Where that left me
Somewhere in the middle of all that dialling back, the caption for the picture arrived on its own:
Turns out neither of us has all the answers. Not the human. Not the machine.
My kid, a human, took a photo that no studio would have staged, in a room that no art director would have chosen, and it worked. The machine built a studio my budget can't reach. And neither one of us could finish the job alone. The machine couldn't keep my eyes. The human couldn't build the room.
The prevailing argument wants me to pick a side. Either these tools are coming to replace us, or they're going to fix everything. You already know how I feel about arguments shaped like that. Extremes are comfortable because they're simple, and simple is what keeps failing us.
I'd rather stay where the work actually happens, which is in between.
To everyone, I say: Love and respect every individual on this planet and avoid extremes. We're all human, and no one possesses all the answers.
Not the machine, either.