https://archive.biennial.com/journal/issue-9/excavating-ai-the-politics-of-images-in-machine-learning-training-sets

Reflection on Excavating AI: The Politics of Labels and the Illusion of Objectivity

1. The Hegemony of Semantics: Classification as Politics

Classification is never a neutral act. Machine learning datasets, such as ImageNet, attempt to force the chaotic complexity of the human world into tidy, predefined boxes. This act of categorization is, in itself, an exercise of power.

Because：

Images are remarkably slippery things, laden with multiple potential meanings, irresolvable questions, and contradictions. Entire subfields of philosophy, art history, and media theory are dedicated to teasing out all the nuances of the unstable relationship between images and meanings.

Images are open to interpretation and reinterpretation.

• Ontological Violence: When developers choose to use WordNet (a hierarchical database of nouns) to organize images, they are effectively asserting that every object, person, and emotion in the world can be reduced to a single noun.
• The Verdict, Not Identification: Within ImageNet’s "Person" category, there are numerous labels like "bad person," "slattern," or "loser"—terms heavy with moral judgment. When an algorithm associates a person’s portrait with these words, it isn't "recognizing" them; it is sentencing them.

2. The Treachery of Images: The False Leap from Appearance to Essence

• The Disconnection of the Symbol: Magritte’s The Treachery of Images proved that "the image of a pipe" is not "the pipe itself." Yet, in machine learning, this logic is distorted. Algorithms operate on the arrogant assumption that by capturing pixel features (the appearance), they can derive the internal essence of the subject.

• The Return of Physiognomy: Crawford and Paglen offer a scathing critique of the resurgence of "phrenology" or "physiognomy" in modern AI. Many training sets attempt to predict "character," "intelligence," or even "criminality" from facial images alone. This creates a dangerous logical death-loop: Because I have labeled you as such, that is who you are.

3. Defining the World: Who Holds the Pen in ImageNet?

The authors reveal a power structure divided into two flawed layers:

The Architects at the Top: Scientists like Fei-Fei Li hold the ultimate power—the design of the Taxonomy.

structure of WordNet

1. The overall taxonomy (the aggregate of classes and their hierarchical nesting);
2. The individual classes (the singular categories images are organized into, e.g., ‘apple’);
3. The specific labeling of each individual image.