The central insight is that faithful translation is impossible, because fidelity would require the preservation of everything that made the original what it was — the context, the history, the social conditions of production, the relationship between speaker and community. Every translation captures something real. Every translation loses something irreplaceable. The good translation is more dangerous than the bad one because the good translation produces the experience of understanding; the reader believes she has received the original when she has received a conversion.
The natural language interface performs a translation that is not from English to code but from the messy, ambiguous, culturally situated thought of the human user into the clean, parseable, culturally deracinated form the model requires. The user experiences this as the machine understanding her. What is actually happening is that the machine is converting her utterance into a form it can process, and the conversion transforms what it carries. The experience of being met is the experience of being translated.
Consider a concrete case. A user in Nairobi types: I need a system that helps women in my community track their savings group contributions. The prompt is clear. The model responds with a technical solution — a database schema, a user interface, an authentication system. But the prompt has already performed a translation the user may not have noticed. The savings group — the chama, in Kenyan usage — is not merely a financial instrument. It is a social institution with specific rules of reciprocity, trust, and mutual obligation that do not map onto the categories of Western financial technology. The model's response addresses the financial function; the social function remains invisible because the model's categories cannot see it.
The Deleuze failure Segal himself describes in You On AI is an instance of translation-betrayal caught by expertise. Claude produced a rhetorically effective but philosophically inaccurate use of Deleuze's concept of smooth space; the passage sounded right because it performed the function of insight, but the reference was wrong in a way obvious to anyone who had read Deleuze. Segal caught it because he is a careful reader. The question the framework presses is how many such betrayals a less specialized user misses — how many times the model produces a culturally situated concept in deracinated form, and the deracination goes unnoticed because the user lacks the context to recognize the conversion.
Spivak's engagement with translation is inseparable from her career. Her 1976 translation of Derrida established her as a major intellectual figure; her subsequent translations of Mahasweta Devi's Bengali fiction (1995, 2002) extended the work into a different register — literary translation from a subaltern language into English for a global audience.
Her 1992 essay The Politics of Translation articulates the theoretical framework most directly: translation is always a political act, always involves asymmetries of power, and always requires the translator to take responsibility for what the conversion transforms. The good translator, in Spivak's formulation, is one who makes her own interventions visible rather than concealing them beneath a surface of fluency.
Fidelity is impossible. Translation cannot preserve everything that made the original what it was, because meaning is not a decontextualizable content but a property of relationships the new context cannot reproduce.
Fluency conceals conversion. The good translation is more dangerous than the bad one because its quality makes the betrayal invisible.
The critical apparatus is essential. Translation without visible accountability for what it transforms is translation that compounds its own violence.
The interface as lens. Natural language interfaces are not windows but lenses — clarifying what they focus on and blurring everything else, producing the experience of directness that conceals the mediation.