Here's a question most AI tool reviews don't ask: what does a photograph actually know?
A photo isn't just pixels arranged to look nice. It's a record of physics. The way light falls across a cheekbone, the shadow pooling under a chin, the gradient of brightness on a wall — all of it encodes real information about a real moment in the real world. And Adobe's newest Photoshop feature just made that information optional.
What happened
Adobe has released a Relight tool in Photoshop that uses AI to re-illuminate photographs after they've been taken. Drag a virtual light source around the image and watch the shadows shift, the highlights move, the entire mood transform — all without going back to the studio.
The reaction has been swift and polarised. Supporters see a retouching breakthrough that could save hours of reshooting. Critics say the results look like video game cutscenes — technically impressive but perceptually off. The word "uncanny" keeps surfacing.
Why the results feel wrong
I think both camps are missing the more interesting point.
The critics aren't wrong — many of the results do look artificial. But the reason isn't that the AI needs better training data or a few more updates. It's that light in a photograph isn't a cosmetic layer. It's structural.
When a photographer sets up lighting, they're not just making things look pretty. They're encoding information. The angle of light tells you something about time and place. The quality of a shadow reveals what a surface is made of. The way light wraps around a face describes its actual three-dimensional form. Our visual systems have been reading these cues for our entire evolutionary history. We don't consciously process them — but we absolutely detect when they're wrong.
And that's where the uncanny quality comes from. It's not a rendering flaw. It's your brain noticing that the highlight on someone's nose doesn't quite match the shadow under their ear, or that the light suggests a source nothing else in the scene agrees with. The AI can move the big, obvious shadows. But a photograph contains thousands of subtle lighting cues — micro-shadows, subsurface scattering in skin, ambient reflections between surfaces — that all need to be perfectly consistent for the result to read as real.
One inconsistency and the whole image starts to feel like a composite. Because that's exactly what it is.
This is different from colour grading
Here's the distinction I think matters. There are two fundamentally different categories of image adjustment.
Surface adjustments change how an image is presented — exposure, contrast, colour grading, sharpening. These are the digital equivalents of darkroom techniques. They work within the physical reality of the photograph.
Structural adjustments change the physical story the image tells. Moving a light source. Altering perspective. Reshaping geometry. These don't enhance the photograph — they fabricate a new one.
We've always been comfortable with the first category. Photographers have been pushing and pulling exposure since the darkroom era, and nobody calls that cheating. But Relight sits squarely in the second category, and I think we need to be honest about that.
The tool itself isn't the problem. For product photography, quick concept mockups, or salvaging a shoot where the lighting was slightly off, the ability to adjust light in post is genuinely useful. The problem comes when structural manipulation gets treated as casually as bumping the contrast. When "just fix it in Relight" replaces the decision a photographer used to make before pressing the shutter.
What this means for designers who work with photography
For photographers, lighting has always been the craft. It's what separates a snapshot from a portrait, what you spend years learning to see and control. Turning it into a post-production slider doesn't democratise that knowledge — it sidesteps it entirely.
And for designers working with photography — choosing hero images, directing shoots, building brand visual systems — this creates a question you'll increasingly need to answer: how far can an image be structurally altered before it stops being a photograph and starts being an illustration?
That's not a technical question. It's a design judgement call. And as AI tools hand us more power to rewrite the physics of an image, we'll need sharper instincts for knowing which properties of a photograph carry meaning and which are genuinely malleable.
Colour temperature? Malleable. Crop? Malleable. The direction the sun was shining when the shutter clicked? That's information. Changing it changes what the image is.
Takeaway
Next time you reach for an AI tool that alters something fundamental about a photograph — the direction of light, the perspective, the geometry — ask yourself one question: am I enhancing this image, or fabricating a new one?
Both are valid choices. But only one of them is still photography. Know which one you're making.