In 1989, sociologist Harry Collins spent years attempting to document how physicists built TEA lasers. Despite detailed written protocols circulating throughout the community, no laboratory successfully constructed a working device from documents alone. Success required sending someone to a functioning lab, watching, trying, failing, and adjusting under the guidance of experienced practitioners.

This finding exposed something fundamental about scientific practice that challenges our textbook image of science as a transparent system of methods and procedures. Much of what scientists actually know cannot be written down, yet this inarticulate knowledge proves essential to producing reliable results.

The philosopher Michael Polanyi captured this phenomenon with a memorable phrase: we know more than we can tell. Understanding this tacit dimension reveals science as a deeply embodied, social, and craft-like practice—not diminishing its rigor, but illuminating why scientific communities function the way they do.

Scientific expertise rests on a foundation of perceptual discriminations and bodily skills that practitioners perform without conscious articulation. A trained pathologist recognizes malignancy in tissue samples through pattern recognition developed over thousands of hours at the microscope, yet when pressed to specify the criteria, often struggles to produce rules that would reliably guide a novice.

This is not intellectual laziness or professional gatekeeping. Polanyi demonstrated that skilled performance involves what he called subsidiary awareness—the integration of countless micro-judgments into a focal act. The chemist who senses when a reaction is proceeding correctly, the field biologist who spots a camouflaged specimen, the physicist who feels when an apparatus is properly aligned: all operate through knowledge structures that resist decomposition.

Ethnographic studies of laboratories by researchers like Karin Knorr-Cetina reveal how experimental work depends on what practitioners call having good hands—a phrase acknowledging that identical protocols yield different results depending on who performs them. The difference lies not in following instructions but in subtle timing, pressure, and responsiveness that develop only through practice.

Recognizing this embodied foundation reframes scientific training. What appears as lengthy apprenticeship is not inefficiency in the transmission of explicit knowledge but necessary cultivation of perceptual and motor capacities that no manual can encode.

Takeaway

Scientific knowledge is not only propositional but perceptual and bodily. The hand and eye of the trained scientist contain information that the tongue cannot speak.

The limits of explicit instruction become visible in what sociologists call the experimenter's regress. When a replication attempt fails, was the original finding wrong, or did the replicator lack the necessary skill? Without independent criteria for competent performance—criteria that themselves require tacit judgment—this question cannot be settled by protocols alone.

Collins's studies of gravitational wave detection, cold fusion controversies, and laser construction all converge on a single pattern: knowledge transfers through human movement, not document transmission. Scientific methods spread along networks of trained bodies, with each generation learning through direct contact with the previous one.

This explains why scientific communities concentrate geographically, why postdoctoral exchanges remain central to careers, and why certain techniques remain localized despite widespread publication. The information needed to perform the work is not absent from papers out of secrecy but because it cannot be represented in the medium of text.

The current replication crisis in several fields partially reflects this reality. When a field scales rapidly without preserving apprenticeship structures, explicit methods sections become insufficient. What looks like a methodological failure is often a failure of social transmission.

Takeaway

Science travels on human legs before it travels on paper. A discipline that forgets this mistakes documentation for knowledge and wonders why its findings refuse to stay put.

The tacit dimension casts a revealing light on ambitions to automate scientific discovery. Machine learning systems can indeed replicate certain pattern-recognition tasks—sometimes surpassing human performance in bounded domains like protein folding or image classification. This suggests that some tacit knowledge can be externalized once sufficient training data exists.

Yet automation succeeds precisely where tacit knowledge has been partially made explicit through the accumulation of labeled examples. The harder problem lies in what Collins calls collective tacit knowledge—the socially embedded judgments about what counts as a relevant question, a meaningful anomaly, or a promising direction. These depend on participation in a form of life, not statistical regularities in data.

An automated system can tell us whether a given tissue sample matches patterns of malignancy in its training set. It cannot, without human framing, decide which tissues warrant examination, when diagnostic categories need revision, or what would count as a discovery worth pursuing. The questions that animate science emerge from communities of practice, not from optimization targets.

This suggests a more productive vision of scientific automation: not replacement of human practitioners but augmentation within collaborative arrangements that preserve the social and embodied matrix from which scientific judgment arises.

Takeaway

Automation captures the codifiable residue of expertise while leaving its living source untouched. The question is not whether machines can do science, but which parts of science were always in the doing.

The tacit dimension does not diminish science; it clarifies what kind of human achievement science actually is. Scientific knowledge is not a disembodied system of propositions but a living tradition sustained by communities who have learned to see, manipulate, and judge in shared ways.

This understanding carries practical implications for how we fund laboratories, structure training, evaluate replications, and integrate emerging technologies. Institutions that treat science as mere information transfer will consistently underperform those that recognize its craft dimensions.

To study science well, we must attend to what practitioners do with their hands and eyes, not only what they write in their papers. The most interesting knowledge may be the knowledge that resists being written at all.