Tacit knowledge, the articulation trap

A philosopher pointed out a simple truth that turns out to be profound: we know more than we can tell. You can ride a bicycle, but you cannot fully explain in words how you keep your balance. You can recognise a face in a crowd instantly, but you could not describe how you do it well enough for someone else to follow. He called this tacit knowledge, the vast store of skill and understanding we possess but cannot put into words. Decades later, when engineers tried to build expert computer systems by interviewing top doctors, mechanics, and other experts to extract their knowledge, they ran into exactly this wall. The best practitioners often could not articulate what they did best. Their skill lived in their hands and instincts, not in explainable rules.

The best person at a task is often not the best at explaining it, because much of real skill is tacit, lived in the doing rather than stored as words. This has a sharp and counterintuitive edge for hiring and judging talent: the smoothest explainer in the room is not necessarily the most capable doer. Some people have polished the skill of describing work without being excellent at the work itself, and some deeply skilled people go quiet when asked to explain what they simply know. Watch what people can do, not only how well they can talk about it, because talking and doing are separate skills, and the interview rewards the wrong one. The same wall has arrived again, in a form almost nobody names. Getting useful work out of an AI system means writing down your standards, your constraints, and what a finished job looks like. People who sit down to do that discover they cannot fully produce it, and tend to read the difficulty as a personal failing rather than as the documented behaviour of a general problem. It is the expert systems bottleneck of the 1980s, met one person at a time. Any system that takes its instructions by asking people to describe how they work inherits the gap between what they do and what they can say. The matching trap for anyone trying to explain their own expertise is the curse of knowledge.

Source: Polanyi, Personal Knowledge, 1958, and The Tacit Dimension, 1966, where the claim that we can know more than we can tell is set out. The expert systems case is the knowledge acquisition bottleneck, generally credited to Edward Feigenbaum and treated at length in Hayes-Roth, Waterman and Lenat, Building Expert Systems, 1983. The closest thing to a controlled test is Berry and Broadbent, On the relationship between task performance and associated verbalizable knowledge, Quarterly Journal of Experimental Psychology, volume 36A, 1984, pages 209 to 231, where practice improved people's control of a simulated sugar factory while their answers to questions about how it worked did not improve, and separately, being told how the system worked improved the answers and not the control.

The Tacit Dimension

Michael Polanyi, 1966

The original argument that the most important knowledge often cannot be spoken.