The disciplinary soul of AI: data annotation and the politics of religious erasure

The process of aligning artificial intelligence relies on human-annotated data, which is derived from vast amounts of human labor in the Global South. Annotators are tasked with sorting human language to create "Ground Truth" for machine learning systems. Fieldwork in Bangladesh involved annotators, including Muslim and Hindu speakers, rating volatile online posts related to religious friction. These workers faced significant pressure and economic vulnerability while performing these tasks, often under conditions reflecting global tech extraction. The process reveals that imposing a singular, secular Western framework onto diverse human experiences involves filtering out local, sacred realities by forcing nuanced human judgments into numerical categories.

The mechanism described demonstrates how imposing a secular, Western epistemology through data tagging acts as a form of intellectual colonialism. The system prioritizes the "niti" (rules) over the lived reality of "nyaya" (justice), creating an alignment based on arithmetic agreement rather than actual human experience. This process is mirrored in Foucault’s concept of pastoral power, where knowledge is extracted through classification, transforming the worker into a confessor who internalizes the system's imposed categories. The core implication is that the drive for AI "alignment" risks erasing minority rationalities by setting a singular, quantifiable standard—a universal human value—that inherently favors the majority worldview, thereby serving as a mechanism for systemic erasure rather than true moral consensus. If alignment necessitates forgetting deeply held, non-quantifiable religious grammars, then the resulting intelligence is not aligned with humanity but with a historically specific, Westernized form of control exerted through algorithmic governance. What responsibility exists in demanding "alignment" when the process itself embodies an asymmetrical power structure?