A photograph that has been subtly, almost courteously, altered by a machine contains a certain kind of silence. Sometimes it goes unnoticed at first. The light appears to be correct. The composition is sound. However, something has escaped the frame, and the picture continues to advance as if nothing had happened. What worries me the most about the current generation of AI image tools is that. They don’t make a big deal out of what they take out. They are neat.
Jessica Smith, an Australian Paralympic swimmer, wrote earlier this year about using an AI to create a self-portrait. She was given two arms by the system. She had one from birth. It is a tiny, yet not insignificant, detail in the architecture of a single image. Throughout her adult life, Smith has told people that her body is not a problem that needs to be fixed. The algorithm made a different decision based on whatever enormous and unexamined pile of images shaped it. She was corrected by it. Silently.
This is the peculiar new realm that generative photography has ventured into. Without being asked, tools that can visualize a boardroom in Singapore or a sunny beach in Lisbon can also smooth away a scar, a wheelchair, a missing limb, or a hearing aid. There is no declaration of absence. It is designed. Additionally, most viewers are unaware of what has been edited out of existence because the output appears plausible.
This was done on purpose by the Deep Angel project, which was run by MIT’s Media Lab years ago. It asked viewers to notice when objects were removed from photos. It was presented as provocation, art, and a means of exposing a mirror to manipulation. The current situation feels different. The experiment is not the erasure. It is the standard behavior. There was no button marked “remove disability.” Long before the user typed a prompt, the training data completed the task.
It’s possible that none of this was intended by the engineers who built these systems. Most of them, I think, didn’t. According to a recent Microsoft article, the issue isn’t intent. It’s the lack of specific bodies, skin tones, or lives in the data that the models are trained on. Trash in, gospel out. A model trained primarily on stock photos of physically fit individuals will confidently create a world composed solely of physically fit individuals. The risky aspect is the confidence.
Photographers have begun to rebel. David Mesfin started a project to retrain models using never-before-seen images after witnessing AI tools repeatedly fail to render Black surfers. Others are discreetly performing comparable work for elderly faces, disabled subjects, and the kinds of beauty that the industry initially neglected to capture on camera. It is labor that is slow and unglamorous. It doesn’t follow any trends.
There’s a sense, watching all of this unfold, that we are deciding what the visual memory of this decade will look like, and we are doing it by accident. The models of the future will be trained using the images we create today. Absence compounds if it is baked in now. A child born next year without a left arm may grow up searching for images that look like her and finding only the corrected version, the one the machine preferred.
It’s difficult to ignore the amount of faith we put in something we never requested to see. The frame expands. The picture shows up. We continue to scroll. A body that once existed has ceased to exist somewhere inside that scroll, and no notification has been sent.
