Built with Microsoft
Artificial intelligence has barely seen us.
Ask most models for an image of a person with albinism and you get something wrong, something medicalised, or something that looks like a ghost. Models learn people from pictures, and we were not in the pictures. The AI Community Library is how we fix that at the source.
Every image in this library was given, not scraped. Consent is the whole design.
Image count and composition to confirm
The problem
A bias you can see in one prompt.
Representation in a training set decides what a model believes a person looks like. Persons with albinism are close to absent from the image data the current generation of models learned from, and it shows.
- Return a white European face when asked for an African person with albinism
- Wash out skin to a flat grey or ghostly white that no human has
- Default to a clinical or pitying frame: the medical photograph, the victim portrait
- Miss the eyes entirely, which is where albinism actually shows
- Fail to render children, elders, professionals, athletes, anyone ordinary
Add before-and-after prompt examples once cleared for publication
- Keep African features and facial structure intact
- Render real skin tone with real variation, freckling and sun texture
- Show hair, eyes and lashes accurately rather than approximating
- Place people in ordinary life: work, school, sport, family
- Represent the full range of ages, genders and body types
Why it matters
This is not a vanity problem.
Image models are already being used to make textbooks, health materials, advertising, public information campaigns and the illustrations in the next generation of children's books. If a model cannot render a person with albinism accurately, that person quietly disappears from all of it, exactly as they have disappeared from every other picture.
It runs deeper than pictures. The same gap shows up wherever computer vision meets a face: photo tagging, camera exposure, identity verification, medical imaging tools trained mostly on pigmented skin. A dataset gap becomes a service that does not work for you.
Fixing it is cheap compared to the alternative. It needs images that exist, given by people who chose to give them, labelled properly and made available to the people building the models. That is all this library is.
- Generated imagery in education, health and media
- Face detection that fails on very low-contrast features
- Camera and phone exposure metering built around pigmented skin
- Dermatology AI trained without depigmented skin in the set
- Stock and search results that return nothing usable
Cite published research for each claim before this page goes live
How it was built
Given, not scraped.
The usual way to build an image dataset is to take pictures off the internet and ask forgiveness later. For a community that has spent decades being photographed without permission, that was never an option.
Invited
We went to our own community first and asked people whether they wanted to be in it.
Consented
Every contributor agreed in writing to how their image may be used, and can withdraw it.
Captured and labelled
Photographed and described so a model learns the right thing from each image.
Released
Made available to AI builders under terms that protect the people in it.
The library today
Where it stands
Add yourself
The library only works if we are in it.
If you have albinism, your photograph teaches a machine what a person with albinism looks like. You choose how it may be used, and you can take it back out at any time.