We See What We Have Learned to See

Perception, like clouds, is shaped more by the observer than the object. One person sees a horse galloping through the sky; another sees a face, a dragon, or nothing at all. The sky doesn’t change—the viewer does. And so it is with images. We do not see the world as it is; we see it as we have learned to see it.

This is not metaphor. It is biology.

Recent research in cognitive neuroscience, particularly by Sylvie Chokron and Christian Marendaz, confirms that vision is an active process of reconstruction. Our brains don’t passively absorb the visual world; they selectively rebuild it using layers of memory, expectation, emotion, and cultural context. Left and right hemispheres parse different spatial frequencies—one favoring detail, the other broad shape and both are deeply shaped by past experience. The visual cortex is not a mirror. It is a storyteller.

Perception as Reconstruction

We like to believe that what we see is what is. But what we see is, at best, what remains after filtering reality through who we are, where we come from, and what we’ve previously seen. Our visual understanding of the world is constructed, shaped by repetition, exposure, emotion, and education.

How We See Photographs

In advertising, this malleability is a feature. The goal is not to depict reality, but to manipulate it, to trigger associations, to frame desire. Stock photography works the same way: it feeds predictable patterns into visual culture so that viewers instantly recognize the feeling or idea being sold.

Photojournalism, on the other hand, operates on the pretense of neutrality. It insists on being evidence, not suggestion. Yet it too is subject to the selective nature of vision. It too is framed, composed, chosen. A photograph may testify to what happened, but the meaning it conveys is always reconstructed in the mind of the viewer. And that meaning is shaped by what the viewer has learned to see.

Enter Generative AI

This is where things get complicated.

Generative AI models, like our perception, build images from what they’ve learned. Their training data is vast, millions of photographs, artworks, captions, and texts. When they create an image, they are not inventing so much as reassembling. Not unlike memory.

And here the lines begin to blur.

If our visual perception is already a reconstruction,a cognitive synthesis of memory and expectation,then what truly separates an image created by a human camera and one created by a machine? Especially when the machine’s “vision” is shaped by our collective visual history?

Some generative images go a step further: they recreate scenes described by witnesses. Trained not on pixels but on recollections. They give visual form to what people remember, not what was recorded. These are not photographs. They are synthetic memories. But they often feel just as real.

We’ve already seen this unfold. In Le jour où j’ai tué mon frère, psychoanalyst Serge Tisseron recounts a haunting experiment: remembering a childhood photograph so vividly he can describe and sketch it—yet when he finally recovers the actual photo, it’s not the same. The details he was certain of weren’t there. The image in his head had been edited by emotion, time, and interpretation.

He then uses generative AI to recreate what he thought he saw. And that version, although artificial, feels more emotionally accurate than the original.

The AI didn’t hallucinate. It collaborated—with memory. The result wasn’t documentation. It was intimacy, stylized. And yet, it remains untethered from what actually occurred.

The Difference That Still Matters

So does it matter?

Yes. Because reality has consequences.

A light-based photograph captures a moment anchored in time and place. However subjective its framing, it retains an indexical link to the real. It may be interpreted, misread, weaponized, but it remains a document of something that happened. A trace.

An AI-generated image, even one based on collective memory, is untethered. It mirrors sentiment, not occurrence. It holds affect, not evidence. It may be more emotionally accurate than a poorly timed snapshot, but it cannot replace it.

And that distinction is not philosophical. It’s ethical. Political. Epistemological.

The danger is not that we will be fooled by AI images. The danger is that we will stop caring whether anything actually happened at all.

A Future of Seeing

In a world where both humans and machines reconstruct the visible from what they’ve learned, the value of the light-captured photograph doesn’t diminish, it sharpens. Because it carries the burden of anchoring us to what occurred, even as our interpretations fragment.

We see what we have learned to see. Machines now do the same. But only one of us was there.

And that still makes all the difference.