Simon Schaffer

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI treats the winter of 2025 as a threshold moment—a before-and-after from which there is no return. Schaffer’s framework complicates this narration without dismissing it. The capabilities that arrived in December 2025 did not materialize from a vacuum. Transformer architectures had been published in 2017. Large language models had been scaling since GPT-2 in 2019. The ability to generate functional code from natural-language descriptions had been demonstrated by GitHub Copilot in 2021 and had improved continuously through 2024. The trajectory was continuous; there was no hidden discontinuity in the technical record. What arrived in the winter of 2025 was a convergence of social conditions that produced recognition—a critical mass of demonstrated capability, accumulated testimony, and interpretive framework that allowed the community to certify that something real had happened. The orange pill changed color not when the technology crossed a threshold but when the community was prepared to see it.

This analysis shifts attention from the technology to the social machinery of recognition—and that shift is both analytically productive and practically consequential. The twenty-fold productivity multiplier in Trivandrum is a fact constructed through a specific instrument within a specific framework of measurement that defines productivity as rapid generation of functional code across multiple domains. A measurement framework calibrated to the instrument’s strengths will inevitably show dramatic gains. The gains are real within that framework. But the framework itself is an artifact of the instrument, and understanding the co-production of instrument and fact is essential to understanding what the multiplier actually measures and what it leaves unmeasured: the depth of architectural knowledge the engineers are or are not building through the accelerated process, the long-term reliability of systems built at unprecedented speed, the organizational knowledge that accumulates through slower, more deliberate development.

Schaffer’s analysis of the invisible hands behind every scientific achievement—Hooke behind Boyle, Kinnebrook behind Maskelyne, the human computers behind the astronomers whose names appeared on the publications—applies with force to the AI moment. The invisible labor that makes the large language model possible—the millions of uncredited creators whose texts formed the training data, the data labelers in Kenya and India earning under two dollars an hour to review content that includes graphic descriptions of violence and abuse—does not appear in the smooth conversational interface through which the model presents its capabilities. The interface performs autonomy. The attribution of that autonomy to “Claude” replicates, at vastly larger scale and with far greater efficiency, the same erasure Schaffer has documented across four centuries of scientific knowledge production.

Origin

Schaffer was born in 1955 and trained in the history and philosophy of science at Cambridge, where he has spent his entire career. His intellectual formation was shaped by the emergence of Science and Technology Studies as a discipline in the 1970s and early 1980s—the period when sociologists like David Bloor and Barry Barnes were arguing, against the prevailing view, that the social study of scientific knowledge was not merely the study of error and deviance but could be applied symmetrically to true and false beliefs alike. The insight that would define Schaffer’s career was that this symmetry extended to the physical apparatus of science itself: instruments, experiments, demonstrations were not neutral windows onto nature but socially organized practices that co-produced the facts they appeared merely to reveal.

The collaboration with Steven Shapin that produced Leviathan and the Air-Pump in 1985 established this argument with a historical specificity that made it impossible to dismiss as mere sociological speculation. Boyle’s air pump was a complex, temperamental machine that frequently leaked, required skilled technicians to operate, and produced results only under conditions that Boyle carefully controlled. The “fact” of the vacuum was not observed directly but inferred through the mediating apparatus of the pump, within a framework of interpretation that Boyle had constructed through previous publications, demonstrations, and the cultivation of a witnessing community whose social standing guaranteed the credibility of their testimony. The instrument and the fact validated each other circularly. That circularity was not a flaw in the method. It was the method.

Subsequent essays developed this framework across the full range of the Scientific Revolution: the Kinnebrook affair and the “personal equation” that revealed the observer’s body as part of the instrument; Newton’s prism experiments and the social processes through which the corpuscular theory achieved dominance over Huygens’s wave theory; Babbage’s calculating engines and the political genealogy of the concept of “machine intelligence” itself. In his 1994 essay “Babbage’s Intelligence,” Schaffer showed that the attribution of intelligence to calculating machines depended on the prior social devaluation of the human computational labor they were designed to replace—a dynamic that, as the AI moment makes visible, has not ceased operating.

Key Ideas

The co-production of instrument and fact. Scientific instruments and the facts they measure validate each other in a circle: the instrument produces results that are interpreted through a framework that the instrument’s prior results helped establish. The thermometer defined temperature as the thing thermometers measure; the telescope defined celestial observation as the thing telescopes make visible. Neither is disqualifying—this is the normal mechanism of scientific knowledge production. But co-production means that claims about what the technology does must be accompanied by analysis of the frameworks through which those claims are produced, because those frameworks are artifacts of the technology, not neutral vantage points from which the technology can be assessed.

Witnessing communities and the politics of demonstration. A finding becomes a discovery when a community certifies it, and the community’s capacity to certify depends on interpretive frameworks, instrumental competences, and institutional structures that must themselves be constructed. Boyle needed gentlemanly witnesses whose social standing guaranteed disinterestedness; Galileo spent years distributing telescopes and training observers before the community capable of seeing Jupiter’s moons existed. The AI moment exhibits the same dynamic: the demonstrations that construct the orange pill—viral posts, revenue curves, the Trivandrum sprint—are performances designed to produce conviction in a witnessing community whose existence was itself produced by the demonstrations.

The politics of intelligence. Schaffer’s 1994 analysis of Babbage identified what he called the politics of intelligence: the recognition that the attribution of intelligence to machines has always been entangled with the devaluation of the human labor the machines replace. A task performed by low-status workers is not considered to require intelligence; when a machine performs the same task, the machine is celebrated as intelligent. The attribution follows the labor hierarchy, not the cognitive complexity. This dynamic operates with remarkable fidelity in the contemporary AI moment: the tasks AI performs most impressively are tasks that, when performed by humans, require skill and creative judgment—but the discourse systematically reframes them as pattern-matching and statistical prediction, positions that make the human version of the work seem less cognitively demanding and the machine’s performance less disruptive than it is.

The controversy study. Scientific controversies reveal the social machinery of knowledge production that settled consensus renders invisible. The controversy study methodology examines disputes while they are active, before one interpretation achieves dominance and is retrospectively naturalized as “what the evidence showed all along.” The AI moment is an active controversy: the competing frameworks—AI as amplifier, AI as pathology, AI as structural displacement—are still contesting, and their contest is being shaped by institutional power structures that determine whose evidence counts and whose is scrolled past. Understanding this contest as a social process, not merely an intellectual debate, is the prerequisite for building constructions honest about their own architecture.

Making up the moment. The concept of “making up” in Schaffer’s usage means not fabrication but constitution: the active construction of a category, a moment, a discovery through the collective labor of a community. The orange pill moment is being made up right now—not in the sense that it is fictional but in the sense that its meaning, its boundaries, and its significance are being actively produced by the community that experiences it. The book that names the moment participates in the construction of the moment. The construction is not neutral. It selects, emphasizes, and privileges certain interpretive frameworks while marginalizing others—and the quality of the construction will determine the quality of what follows.