The book emerged from Latour's unusual intellectual trajectory. Trained in philosophy and theology in France, he took up anthropological fieldwork in Côte d'Ivoire before arriving at the Salk Institute in 1975. His plan was to apply ethnographic methods — the kind used to study 'exotic' tribes — to a Western scientific laboratory. The resulting observations were produced from a position of productive naivety: Latour did not know the biochemistry, and his ignorance allowed him to notice what participants took for granted.
What he noticed was that the laboratory was saturated with inscription devices — instruments that transformed biological phenomena into traces, charts, graphs, and eventually published papers. The fact that Guillemin eventually won the Nobel Prize for, the chemical structure of TRH, existed nowhere in nature in a directly accessible form. It existed as the endpoint of a long chain of inscriptions, each of which transformed the previous step, and the final inscription — the published structure — depended on every preceding link. The 'fact' was a network achievement, stabilized through publication and institutional recognition.
The book's argument provoked controversy. Scientific realists read it as a denial that science produces knowledge of reality — as a reduction of scientific claims to social construction. Latour and Woolgar's more careful position was that scientific facts are real (they refer to features of the world) and constructed (they are produced through concrete chains of practice), and that the dichotomy between 'real' and 'constructed' misses the actual operation of science. The subtitle of the second edition — 'The Construction of Scientific Facts' — became a flashpoint; Latour himself later said he regretted the word 'construction' because of the way it was misread.
For AI, the book provides both method and concept. The method — patient ethnographic tracing of concrete practice, refusal to accept participants' self-descriptions at face value, attention to instruments and inscriptions — is what any serious study of AI networks requires. The concept — the fact as network achievement, the inscription as transformative mediator, the publication as stabilization event — maps with remarkable precision onto the practices of AI training, deployment, and output. Claude's outputs are not 'discovered' features of natural language; they are inscriptions produced through a chain of training, reinforcement, and interaction that is every bit as concrete and network-embedded as Guillemin's laboratory work.
The book was based on Latour's 1975–1977 fieldwork at the Salk Institute, supported by a Fulbright fellowship. Steve Woolgar, a British sociologist, co-authored the analysis. The first edition was published in 1979; a significantly revised second edition appeared in 1986 with a new postscript addressing the criticisms the book had generated.
The book was the founding text of a research program — sometimes called 'the sociology of scientific knowledge' or 'science and technology studies' — that produced dozens of subsequent laboratory ethnographies across the 1980s and 1990s. The program transformed the study of science and has had widespread influence on how scholars understand the production of knowledge in technical fields.
Ethnography of the laboratory. Apply anthropological methods to scientific practice. Observe without accepting the participants' self-descriptions as analytical conclusions.
Inscription devices. Instruments that transform phenomena into traces, charts, and texts. The output of every inscription device becomes the input to the next, producing chains of transformation.
Facts as network achievements. Scientific facts are produced through concrete chains of practice. They are real, and they are assembled; the two characterizations are compatible.
Publication as stabilization. The peer-reviewed article performs the final stabilization that converts a working claim into a citable fact. Once stabilized, the chain of inscription that produced it becomes a black box.
Productive naivety. The ethnographer who does not know the technical content can see what the participants take for granted. This is the observational position the analysis of AI networks also requires.