Trained Judgment (Daston–Galison)

In the AI Story

Trained judgment differs from truth-to-nature in a crucial respect: it does not claim that the expert sees the ideal form behind the accidents. It claims, more modestly, that the expert has developed a tacit repertoire of recognition — an internalized understanding of what artifacts typically look like, what normal variation encompasses, what unusual findings should prompt further investigation. The authority is neither in the Platonic form nor in the mechanical causation but in the pattern-recognition capacity that only sustained practice in a specific domain produces.

The regime's institutional infrastructure is extensive and specific. It includes training programs that deliberately expose novices to characteristic errors of the instruments they will use, atlases that pair mechanical reproductions with expert annotations indicating what should and should not be trusted, professional communities that transmit tacit standards through apprenticeship rather than codification. Trained judgment is an institutional achievement as much as a cognitive one: it exists only where communities sustain the practices that produce and transmit it.

The regime's confidence artifact is demonstrated fluency — the ability of the expert to move through a domain's characteristic problems with the particular quality of responsiveness that reveals deep familiarity. This confidence artifact is genuinely reliable within its domain and genuinely difficult to fake outside it, because the tacit knowledge it reflects cannot be acquired without the sustained engagement that produces it. Until recently, this constraint held. The signal was reliable because counterfeiting it required acquiring the very competence the signal indicated.

The AI transition has partially broken this constraint. Large language models can reproduce the surface features of trained judgment — the fluent handling of technical vocabulary, the appropriate qualification of claims, the recognizable rhetorical patterns of expert communication — without the domain immersion that has historically produced these features. The decorrelation is the central problem this volume addresses, and it marks trained judgment's characteristic confidence artifact as the specific heuristic most vulnerable to AI-generated simulation.

Origin

Daston and Galison introduced trained judgment in Objectivity (2007) as the third of three successive regimes of scientific representation. They located its emergence in specific twentieth-century developments — particularly in medical imaging, atomic physics, and neuroscience — where practitioners discovered that mechanical reproduction alone could not support reliable research. The framework was refined in subsequent work on the history of scientific training, expert witness testimony, and the institutional structures of professional communities.

The concept drew on and extended Michael Polanyi's work on tacit knowledge, Ludwik Fleck's analysis of thought styles, and Harry Collins's ethnographic research on scientific communities. What Daston and Galison added was the historical specificity: trained judgment was not a universal feature of expertise but a particular epistemic regime with identifiable origins, characteristic institutional forms, and its own vulnerabilities to technological disruption.

Key Ideas

Judgment rehabilitated. The expert's interpretive capacity is not a distortion to be suppressed but a trained resource to be cultivated through sustained domain engagement.

Tacit repertoires over explicit rules. Expertise consists of pattern-recognition capacities that cannot be fully articulated and must be transmitted through apprenticeship-based practices.

Institutional infrastructure is constitutive. Trained judgment exists only where communities sustain the training programs, annotated atlases, and professional norms that produce and transmit it.

Demonstrated fluency as confidence artifact. The expert's characteristic responsiveness to a domain's problems becomes the surface signal of reliability — historically hard to counterfeit, now mimicable by statistical systems.

The specific vulnerability to AI. Large language models produce the surface features of trained judgment without the tacit substance, breaking the correlation that made the signal reliable.

Debates & Critiques

One sustained debate concerns whether trained judgment is a genuinely novel regime or a rehabilitation of truth-to-nature under different vocabulary. Critics argue that Daston and Galison overstate the novelty; defenders respond that the specific institutional infrastructure of twentieth-century trained judgment — the programs, the annotated atlases, the professional accountabilities — differs in kind from the eighteenth-century apprenticeship of botanical illustration. A related debate concerns whether the current AI moment is undermining trained judgment as an epistemic regime or merely exposing its always-present dependence on institutional scaffolding that the digital environment has weakened.