Can the Subaltern Prompt?

In the AI Story

The question is not hypothetical. It describes the condition of the majority of the world's population. A farmer in rural Bihar possesses knowledge of soil composition, seasonal variation, crop rotation, water management, and sustainable agriculture accumulated over generations of practice, observation, and communal transmission. This knowledge is real, testable, and in many cases outperforms Western industrial agriculture on sustainability metrics. It is different — organized by different categories, transmitted through different media, validated by different criteria. The farmer cannot prompt. Not because she lacks intelligence, creativity, or will to build, but because prompting requires three things she does not have: fluency in the language the model understands best, access to the conceptual categories the model recognizes as knowledge, and the infrastructure to reach the model at all.

Each barrier reveals something about the architecture of exclusion. The language barrier is quantifiable: English accounts for over half of most major training corpora, the top ten languages account for more than ninety percent, and the remaining six thousand-plus languages share the fragments. The conceptual barrier runs deeper: the model does not merely prefer English; it prefers the epistemological categories that English-language academic and technical discourse has produced — propositional claims, universalist aspirations, textually archived evidence. The infrastructure barrier is material: hardware costs more relative to local wages in Lagos than in San Francisco, connectivity is intermittent and expensive, and the assumption that users have continuous high-bandwidth access is baked into the interaction design.

Consider what happens when the farmer, through some combination of access and translation, does manage to prompt. She asks about soil management for her specific conditions — alluvial soil in the Gangetic plain, monsoon variability, small-plot farming in a region where the water table has been dropping for decades. The model answers fluently. The answer may even be useful. But the answer is structured by the categories of Western agricultural science: it frames her situation as a problem to be solved rather than a relationship to be maintained, recommends interventions rather than acknowledges practices, cites studies conducted in experimental stations rather than knowledge accumulated over centuries of situated practice. The farmer faces a choice that is also a loss: adopt the model's framing and gain access to its recommendations, or maintain her own framing and remain outside the conversation about the future of agriculture.

The Orange Pill celebrates prompt engineering as the new literacy — the skill that separates those who direct the amplifier from those who are directed by it. Segal's argument that the question becomes the product is powerful. But prompting is itself a form of epistemic gatekeeping. The good prompt is one that speaks the model's language, and the person who can formulate it has already translated her intention into the form the amplifier requires. Those who cannot perform this translation are excluded not merely from the tool but from the future the tool is building — because the future AI constructs is shaped by the questions it is asked, and the questions it can be asked are those askable in its language, its categories, its epistemological frame.

Origin

The formulation transposes Spivak's 1988 Can the Subaltern Speak? to the AI context, preserving the structural form of the original question. The original asked whether the institutional apparatus of knowledge production could receive subaltern speech as meaningful. The transposition asks whether the computational apparatus of AI can receive subaltern articulation as meaningful prompt.

The framing has been developed across a growing literature at the intersection of postcolonial theory and AI studies, including the work of Shakir Mohamed and Marie-Therese Png on decolonial AI and the AI Decolonial Manyfesto drafted by the Decolonial AI Manyfesto Collective.

Key Ideas

The prompt as privilege. The ability to prompt effectively is itself a form of cultural capital, distributed unevenly across the populations the technology claims to serve.

Three barriers, one architecture. Language, conceptual, and infrastructural barriers are not separate problems but manifestations of a single architectural fact: the system was built with a specific user in mind.

The translation loss. Even when the subaltern user does prompt, her intention must be translated into categories the model recognizes, and the translation strips the relational, embodied, context-specific knowledge that made the intention meaningful.

Democratization as integration. Extending access to the system is not the same as building a system that includes; the extension is integration on terms set by the center.

Debates & Critiques

The framework has been criticized for appearing to deny the real gains that AI tools provide to developers in the Global South — gains The Orange Pill documents with evident sincerity. The response, consistent with Spivak's methodology, is that acknowledging the real gains and naming the structural partiality are not contradictory tasks but complementary ones. The gains are real; the gains are partial; the partiality is structural; the structure is invisible from the position of the beneficiary. All four claims must be held simultaneously.