Excavating AI: The Politics of Images in Machine Learning Training Sets
Reflecting on this reading is very important to me, especially when it comes to artificial intelligence and how images are labeled. A lot of the time, we don’t really stop and think about who is labeling these images and how much power that actually has. But this reading made me realize that labeling images is not just a technical thing: it’s a human thing, and it comes with opinions, culture, and bias.
One important question we need to ask is: who is doing the labeling? Humans are very diverse, and the way we see images depends a lot on our background and our culture. For example, when we think about marriage, there is usually a specific image that comes to mind. For some people, it’s a white dress and a big wedding. For others, it’s traditional clothing, different ceremonies, and different cultural practices. None of these are wrong — they are just different ways of seeing the same idea.
But the problem is that when images are labeled for AI systems, it’s often only one group of people deciding what those images mean. In most cases, the people labeling these images are men, and mostly white men. There are very few women involved, and even fewer people of color, especially Black people. That means many cultures, especially cultures from Africa, are not included at all. Their perspectives, traditions, and ways of seeing the world are missing from these systems.
This is dangerous because AI systems are starting to shape how the world works. If AI is trained only on the perspective of a small group of people, then it will reflect only their views. But the world is much bigger than that. The world is diverse. There are many cultures, many histories, and many ways of understanding images and meanings.
The reading also made me think about language. Many languages in Africa were passed down through oral history, not books. Because of that, there is not a lot of written data for AI systems to learn from. Meanwhile, languages like English or French have tons of books and written material, so AI systems understand them much better. Languages from places like Rwanda or other African countries are often ignored, not because they are less important, but because the data was never collected.
So I believe it’s really important that we pause and think about what kind of world we are building with AI. If we don’t include diverse people, cultures, images, and languages, then AI systems will only tell the stories of the people who labeled the data in the first place. That means many voices will be left out.
If we want AI to be fair and useful for everyone, we need to include more diversity — in who labels images, in what data is collected, and in whose cultures and languages are represented. Only then can we build AI systems that truly reflect the world we live in, not just a small part of it.
My Gesture Recognition Project
For this project, I made a simple system that can understand hand gestures using a camera. The main idea is to help people understand gesturesr. By only looking at my hand gestures, the computer can tell what I am trying to say.