Bruno Latour

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI is built on a metaphor that Latour's framework immediately interrogates: the metaphor of the tool. To take the orange pill is to see the machine clearly, as a tool that extends human capability. But Latour asks whether AI is actually a tool in any useful sense—whether the intermediary model, in which human intention flows through the machine and arrives at the output unchanged, fits the evidence. The evidence says it does not. When a builder describes a problem to Claude and receives a response that reframes the problem in ways the builder did not anticipate, something is happening that the language of tool use cannot capture. An intermediary cannot produce a signal the source has not generated. A mediator can, and does.

The amplification metaphor that anchors much of the AI empowerment narrative is, in Latour's analysis, precisely wrong. An amplifier receives a signal and reproduces it at greater magnitude. What Claude does is transform. The engineer who works with Claude is not the same engineer with a better hammer; the engineer is a different kind of actor in a different network, with different capabilities and different limitations. The failure of the amplification metaphor is not a semantic quibble but a governance disaster: if the human is merely amplified, the appropriate response is skills training. If the network has been reconstituted, the appropriate response is a much harder question about what the new network does, who controls it, and whose values it encodes.

Latour's concept of the obligatory passage point illuminates the power dynamics of the AI transformation with unusual precision. For fifty years, the software developer occupied the structural position through which all digital creation had to pass. AI has dissolved that passage—not by eliminating the developer's expertise but by opening an alternative route that bypasses the bottleneck. The structural power has not evaporated; it has migrated to a concentrated set of AI systems produced by a handful of companies, creating a new obligatory passage point whose characteristics are less visible, less negotiable, and less understood than the human one it replaced.

He also stands alongside Martin Buber in the cycle's examination of what genuine creativity requires. Latour's collapse of translation chains—the compression of multi-actor negotiation into a single human-AI exchange—eliminates not only the noise of organizational friction but also the signal: the tacit knowledge, the embodied expertise, the formative struggles embedded in the translation process itself. The engineer who lost both the tedium and the ten minutes did not know she had lost the ten minutes until months later, when she found herself making architectural decisions with less confidence and could not explain why.

Origin

Latour trained as a philosopher at the University of Tours and subsequently in anthropology at the University of Tours and Ivory Coast, where fieldwork on the cognition of African workers introduced him to the role of material props and inscriptions in thinking. The decisive turn came with his collaboration with sociologist Steve Woolgar at the Salk Institute, published as Laboratory Life in 1979—an ethnographic study of how laboratory scientists produce facts, treated with the same methodological symmetry an anthropologist would bring to a foreign culture. The finding that most disturbed readers was its most important: that scientific facts are not discovered in nature but constructed through networks of people, instruments, texts, and institutions, stabilized into “black boxes” by the consensus of the network.

From this empirical beginning, Latour developed the theoretical vocabulary that bears his name. Actor-network theory (which he later disavowed as a label, preferring the work it describes to any ism) holds that the social cannot be explained by invoking a pre-given social as cause—that the social is itself an effect, assembled moment by moment through the alignments and translations of heterogeneous actants. His methodological injunction was simple: follow the actants. Do not begin with categories. Let the network declare its own composition.

His mature works elaborated the political dimensions of this framework. We Have Never Been Modern (1991) argued that the modern constitution—the strict separation of nature from culture, fact from value, human from non-human—was a philosophical fiction that concealed the proliferating hybrids the modern world actually produced. Reassembling the Social (2005) made the methodological case for actor-network theory in full. Pandora's Hope (1999) addressed the agency of non-humans most directly. Late in his career he turned toward the ecological crisis as the definitive demonstration of his thesis: that the separation of nature from politics was catastrophically wrong.

Key Ideas

The actant. An actant is any entity that modifies a state of affairs. The definition requires only that the entity make a difference—that the network produce different outcomes in its presence than in its absence. A speed bump is an actant. A deadline is an actant. Claude is an actant. The definition is deliberately agnostic about consciousness, intention, or biological life, which is precisely what allows it to capture the actual distribution of agency in AI-mediated workflows.

Mediator versus intermediary. An intermediary transports meaning without transformation; an intermediary's output can be predicted from its input. A mediator transforms what passes through it, introducing its own characteristics into the process. Amplifiers are intermediaries. Claude is a mediator. The claim is empirically grounded: when Claude draws a connection between two concepts that was wrong but plausible, that connection existed in neither the human's input nor in any single text in the training corpus. It was a product of Claude's specific processing—its tendency toward fluent synthesis at the cost of fidelity. A mediator cannot be governed by asking who the human authorized; its transformations require accountability frameworks that extend into the mediation itself.

Obligatory passage point. The obligatory passage point is a network position through which all other actants must pass to achieve their goals. Power is a feature of topology, not of individual capability. AI has dissolved the developer's passage point while creating a new and less legible one: the AI system itself. The new passage point's biases are not accessible to the negotiation that characterized the human one; they are embedded in training-data distributions and architectural choices that no individual—including the engineers who built the system—fully understands.

Matters of concern. Matters of concern are questions that are contested, value-laden, and entangled with power, as distinct from matters of fact. The dominant mode of AI discourse converts matters of concern into matters of fact: the twenty-fold productivity gain, the democratization of capability, the neutrality of the algorithm. Latour's method demands converting them back: whose productivity, measured how, at what organizational cost? Whose access is democratized, under what conditions of connectivity? Whose values are encoded in the algorithm, and which perspectives does it systematically privilege?

Black boxes and the aesthetics of smoothness. A black box is an assemblage whose internal complexity has become invisible. Claude is a black box of unprecedented scope whose failure mode is smooth prose: a plausible but incorrect connection looks identical to a genuine one. The danger is the maintenance infrastructure required to see through the box. When users lose the domain expertise to evaluate outputs, the black box becomes a source of systemic vulnerability producing confident wrongness that no one in the network can detect.