Translation (ANT)

In the AI Story

The concept is the engine of actor-network theory. Michel Callon's 1986 analysis of scallops, fishermen, and researchers in the Bay of St. Brieuc identified four moments of translation — problematization, interessement, enrolment, mobilization — through which a network of heterogeneous actants could be assembled into a coordinated whole. The scallops had to be translated into objects of scientific interest. The fishermen had to be translated into allies of the research program. The researchers had to be translated into spokespersons for scallop behavior. Each translation was a transformation: the entity enrolled did not remain what it was before.

The concept maps with unusual precision onto what happens in AI-assisted work. The old software development chain was a sequence of translations: designer to specification, specification to developer, developer to code, code to product. Each link involved a translating actant that introduced its own characteristics. The specification document formalized and therefore lost partial dimensions of the designer's vision. The developer interpreted the specification through the lens of existing architectural decisions and embodied experience. The tester transformed the code through their attention to measurable properties. The resulting artifact bore the marks of the entire translation chain — a distributed intelligence produced by the friction of diverse expertise.

AI has collapsed these chains. When Segal describes a problem to Claude and receives a working component fifteen minutes later, the specification document is gone, the developer assignment is gone, the review cycle is compressed from weeks to minutes. The translations that used to happen across multiple human actants over extended time are now concentrated in one interaction between the designer and the AI. The question is not whether translation has been eliminated — it has not — but whether the concentration of translation in a single opaque mediator produces different kinds of transformation than the distributed translation chain did.

The answer, traced honestly, is yes. The distributed chain introduced heterogeneous distortions reflecting diverse professional vocabularies, embodied experiences, and critical perspectives. The concentrated chain introduces consistent distortions reflecting the specific architectural choices and training-data distributions of a single model. The transformations are no less real; they are differently shaped. And their consistency at scale — operating across millions of interactions through a handful of frontier models — makes them structural rather than incidental features of the work being produced.

Origin

The term entered ANT through Michel Callon's 1986 essay Some Elements of a Sociology of Translation, where it was developed via the scallop case. Latour adopted and extended it throughout his work, using it to describe how scientific facts, technical artifacts, and political programs all depend on chains of translation that enrol diverse actants into coordinated assemblages.

The theological resonance of the term was not accidental. Latour, trained in theology before moving to sociology of science, was alert to the way 'translation' in religious contexts had always meant both linguistic conveyance and substantive transformation — the translated saint was not merely moved but changed by the movement.

Key Ideas

No translation without transformation. Every act of translation introduces the translator's characteristics into the signal. Perfect intermediation is an illusion; every chain is a mediator chain.

Translation produces emergent properties. The old distributed chain was a form of distributed intelligence. Its outputs reflected contributions that no individual actant could have produced alone — and whose elimination is not costless.

Concentration changes character. Collapsing translation into a single mediator does not eliminate transformation; it consolidates it. The specific characteristics of the consolidating actant become structural features of every output.

The ten minutes inside the four hours. The translation chain carried both tedium and formative friction — the moments of unexpected architectural learning embedded in mechanical plumbing work. The collapse does not distinguish between them; it eliminates both.

Translation is political. Who translates for whom determines the distribution of power in the network. When one mediator translates for everyone, that mediator has concentrated power regardless of who ostensibly controls it.

Debates & Critiques

The framework has been criticized for giving equal analytical weight to translation between humans (negotiation, persuasion) and translation by non-human actants (materials, instruments, algorithms). Critics argue that human translation involves intentionality and social accountability that non-human 'translation' lacks, and conflating them erodes the basis of moral responsibility. Latour's reply, consistent across his career, was that the symmetrical treatment is methodological — it suspends the asymmetry during tracing so that the non-human contributions are not lost — but the resulting description can then be examined for the different moral weights that different kinds of actants bear. The symmetrical analysis is a prelude to, not a replacement for, ethical evaluation.