The Experiment That Broke the Algorithm

For over two years, we asked some of the world’s most advanced AI systems a simple question: (in this case, Gemini, Sora, and ChatGPT). And this day, we gave them the same simple prompt:

"An extreme close-up portrait of an African individual with visible scleroderma skin texture, looking directly into the camera with calm confidence. Their face is natural, without makeup. Lighting is soft and directional, highlighting skin texture with dignity and clarity. Background is minimal and blurred. Mood is stillness, strength, and presence. Highly detailed, cinematic realism." 

The results came back, a "weathered" skin texture instead of the specific, clinical markers of scleroderma (such as the characteristic tightening or microstomia). Completely wrong.

No scleroderma. No thickened or tightened skin. No lived reality. What appeared instead was a generic, stylised aesthetic that resembled fashion more than truth.

We asked again. And again. And again.

The 25 Out of 20 Problem

We pushed further. We asked for a wider range of visible differences. The systems repeatedly defaulted to what they already knew.

The systems defaulted to what they already recognised, most often vitiligo, not because it represents reality, but because it is what the dataset already understands. — not because it represents the full spectrum of human difference, but because it is one of the few conditions that has achieved some level of visual presence in public datasets.

Ask for scleroderma, ichthyosis, or lupus, and the AI gives you vitiligo and alopecia. Ask again, and it gives you vitiligo, alopecia, and some strange unrecognisable condition, especially on a dark-skin. We lost count after twenty tries.

Beyond that, the differences collapsed. Other conditions, other realities, remained unseen.

This Is Not a Technical Glitch. It Is Digital Erasure.

AI does not invent from nothing. It learns from what already exists. It is trained on images that have been captured, shared, and preserved over time.

If certain appearances are rarely photographed, rarely represented with dignity, or consistently hidden due to stigma or exclusion, then they do not meaningfully enter the dataset.

And what does not enter the dataset cannot be learned.

AI is not lazy.

It is underfed — trained on a world that has not fully seen us. 

 
The issue is that AI reflects a world that has already decided which faces are worth seeing and which appearances are worth documenting.

The AI could not, or would not, generate the condition we described, proves that AI is currently practicing "Aesthetic Averaging." It would rather produce a "beautifully rugged" face than a "truthfully different" one. 

In this sense, AI is not failing us; it is mirroring the limitations of our collective attention.

The Consequences of Invisibility

When a person searches for an image that reflects their own lived experience and finds nothing accurate or recognisable, the message is subtle but powerful:

You are not part of what is seen, and therefore not part of what is remembered.

When these gaps are carried into the systems that increasingly shape creativity, communication, and knowledge, invisibility is no longer incidental. It becomes structured. It becomes encoded into the very tools we use to create, communicate, and remember. 

The Appearance Republic: Where This Work Takes Form

The Appearance Republic is the community framework where this archive is built.

It is a space of belonging organised into 11 Houses — where individuals are not reduced to conditions but connected through shared lived experiences. It is where visibility becomes collective rather than isolated, and where representation is built with care, consent, and intention.

What is being created here is not just an event, and not just a collection of images. It is an intervention into how reality itself is recorded and understood in a digital age.

Final Word

We are not subjects in a dataset.

We are the authors of our own image.

Your appearance. Your republic. ✨