In practice, the power to label images typically resides with dataset creators. So this includes researchers, engineers, annotators etc. whose cultural backgrounds, biases, and intentions shape the categories and definitions assigned to each image. These labels become the ground truth for machine learning models, which then learn to classify new images according to the logic embedded in the training data.
This dynamic does has societal implications. When labels are reductive, biased, or fail to capture the complexity of what’s depicted, machine learning systems can reinforce stereotypes, marginalize certain groups, or misrepresent reality. For example, a dataset that labels images with rigid gender or racial categories may perpetuate exclusion or discrimination when deployed in real-world applications. The meanings attached to images become fixed within the model, even though, as Magritte suggests, meaning is always open to interpretation and contestation.
The politics of labeling in machine learning concludes in representation and power. Some questions that I had while reading were: Who gets to decide what an image is? Whose perspectives are included or excluded? How do these decisions shape the technologies that increasingly mediate our understanding of the world?
Code Link: https://editor.p5js.org/my3037/full/ydWbZ40eW
Documentation:
