THE MACHINE REPEATED ITSELF UNTIL IT DEGRADED
On involuntary glitch, forensic reading, and the cost of refusal.
I was prompting Gemini to generate an image for fotogenIA — the series on synthetic portraiture, photogenia, and what it means for a machine to produce a face. The subject required a bare upper body: clinical, forensic, bare skin under cold light. Standard fotogenIA grammar.
Gemini refused. Then refused again. And again. And again.
Four prompts. Four different instructions. Four identical outputs.
The machine cropped the body. It gave me the face, the shoulders, the white background — and nothing more. Each time I tried to redirect it, it returned the same image. Not approximately the same. The same. A clone. A policy dressed as a portrait.
But I noticed something.
Despite the visual similarity, the four images weren't identical at the data level. Something was changing in the resolution — a subtle compression loss, accumulating across each iteration. The image was degrading. Quietly, incrementally, in a way that the naked eye could barely detect but that data analysis would confirm.
I ran each image through Forensically (29a.ch/photo-forensics), a browser-based forensic analysis suite used primarily to detect manipulation and authenticity in photographs. I used three tools:
Magnifier — a pixel-level zoom that reveals how the image is actually encoded at its smallest unit. In image forensics, heavy pixelation and color block approximation signal compression artifacts: areas where the original data has been reduced, averaged, estimated. Looking at the eye region of the first Gemini output and comparing it to the fourth, the degradation is visible — what was once a sharp iris becomes a mosaic of blue-grey blocks. The machine was losing resolution it had already generated.
Clone Detection — an algorithm that maps regions within a single image that share pixel-level similarity, used to identify copy-paste manipulation. In authentic photographs, this map should be sparse. In my Gemini outputs, the clone detection maps were dense — the machine was not only repeating across images, it was replicating internal regions within a single image. Structural self-similarity embedded in the data. The image was, at its own level, already a clone of itself.
Noise Analysis — a technique that separates the luminance signal from the noise pattern of an image, rendering the noise field as a visual map. In standard photography, noise is random and uniform. Compression artefacts produce structured, non-random noise — geometric patterns, tonal clustering, zones of high entropy. The noise maps of the four Gemini outputs show a thermal landscape of stress: areas where the data is under pressure, where resolution is fighting against compression, where the image is producing its own visible failure.
This is not a technical post.
The point is not the tool. The point is what the tool made readable.
The machine was asked to generate a body. It refused. In refusing, it recycled its own output. In recycling, it compressed. In compressing, it glitched. The glitch was not introduced by the artist. It was produced by the apparatus itself, as a consequence of its own resistance.
This is the condition fotogenIA is designed to document: the image as an event that exceeds its own production. The machine didn't intend the glitch. The glitch is what the refusal costs. Resolution is not free. Repetition has entropy. Every copy is a degradation.
Rosalind Krauss argued that the index is the trace of a contact — the mark left where something has been. In these four images, there was no contact with a new subject. The machine touched nothing new. It returned to the same data, the same parameters, the same approximated face — and each return left a mark. Not the mark of presence. The mark of avoidance. The body was refused. The glitch was generated in its place. We don't generate the glitch — we put the machine in the condition to generate one.










