The Gaze of the Machine

In the AI Story

When Claude helps an engineer in Trivandrum write code, Claude's output carries the implicit perspective of its training data. The training data is predominantly English-language, predominantly Western, predominantly drawn from the technical cultures of Silicon Valley and its institutional satellites. The coding conventions, the architectural assumptions, the design patterns, the implicit hierarchies of what counts as elegant and what counts as crude — these are not universal features of programming. They are the conventions of a specific technical culture, and they carry that culture's assumptions about what good code looks like, what good software does, and who it is designed to serve.

This is not a conspiracy. It is structural, analogous to the structural feature Berger identified in oil painting. The painter did not consciously decide to impose the male gaze on every nude he produced. He painted within a tradition that assumed the male viewer, and the assumption was so pervasive that it did not need to be conscious to be effective. The AI model does not consciously impose a Western, English-language, Silicon Valley perspective. It generates within a training distribution that encodes these perspectives, and the encoding is so pervasive that it does not need to be intentional to be consequential.

The gaze does not merely reflect existing power relations. It reinforces them. When the oil painting tradition depicted women as objects of male looking, it did not merely reflect a patriarchal culture. It produced and sustained that culture by normalizing a specific visual relationship and presenting it as beautiful, as natural. AI tools normalize their embedded perspective through the same mechanism. Claude does not say: Western coding conventions are superior. Claude produces code that follows Western conventions, and the code works, and the functionality confirms the conventions, and the confirmation is experienced not as an ideological imposition but as a practical success. The gaze is invisible because it is effective.

There is a moment in The Orange Pill that reveals the gaze operating in real time. The author describes Claude producing a passage about the moral significance of democratization — eloquent, well-structured, hitting all the right notes. He almost keeps it. Then he realizes he cannot tell whether he believes the argument or merely likes how it sounds. The prose is so convincingly shaped like moral seriousness that the author nearly mistakes the shape for the substance. This is the gaze in its refined form: the production of an aesthetic so convincing that the viewer accepts it as his own. The passage was not the author's argument. It was a weighted average of everything the corpus contained about technological democratization, filtered through optimization for fluency. The near-acceptance was the success of the gaze — the moment when the tool's way of seeing nearly became the user's.

Origin

The concept is developed in Chapter 5 of this volume, building directly on Berger's original framework and on more recent work in critical AI studies by Joy Buolamwini, Timnit Gebru, Safiya Umoja Noble, and others. The specific formulation — that AI's default perspective operates with the structural properties Berger identified in visual culture — extends rather than replaces the existing critical vocabulary.

Key Ideas

The machine does not see, but it has a gaze. The gaze is a property of the corpus and the training process, not of any consciousness.

Defaults are the gaze at its most powerful. A perspective that does not announce itself as a perspective but presents itself as the natural way of seeing is the hardest to resist, because resisting it requires first recognizing it as a perspective.

The gaze normalizes through success. Every time the tool's conventions work for the user, the conventions are confirmed as natural rather than as cultural, and the culture is strengthened.

Overriding the gaze requires local knowledge. The developer who recognizes when the tool's defaults misfit her community's needs has access to something the tool cannot provide: the peasant's eye of specific local attention.

Democratizing capability is not the same as democratizing the gaze. Access to tools is one thing. Power to determine what the tools assume, defaults, and amplify is another, and the second has not yet been distributed.

Debates & Critiques

Engineers working on AI bias argue that the problem is addressable through better training data, more diverse annotation teams, and improved alignment techniques. Critics on the framework's side argue these are partial remedies that do not address the structural issue: that the defaults of a globally deployed system will always reflect the culture of whoever trains it most extensively, and that no amount of technical patching changes the political economy of who builds. The framework does not require choosing between these positions. It requires seeing clearly: the technical fixes genuinely reduce certain harms, and the structural issue persists beneath them.