A baker can tell when a dough is ready by pressing two fingers into it and watching how slowly the dent fills back. The measurement exists nowhere on paper. It is not a temperature or a time. It is a thing the hand has learned across thousands of repetitions, and if you asked the baker to explain it precisely, the explanation would fall short of what they actually do. They know more than they can say.
The philosopher Michael Polanyi gave this its clearest formulation: we know more than we can tell. He was interested in the kind of knowledge that resists being made explicit, recognising a face, riding a bicycle, judging the moment a sauce splits. None of it lives in instructions. It lives in the body, accumulated, mostly unconscious. He called it tacit knowledge, and once you have the term, you start seeing it everywhere skilled work happens.
It is worth being careful here. This is not an argument against machines. Automation has done real good. It has taken work that broke people’s bodies and made it survivable, taken processes that were dangerous and made them dull. To romanticise the hand at the expense of the engine is to forget how much hand-labour was misery. The question is narrower and more interesting than for-or-against. When a process becomes industrial, what exactly gets carried over, and what gets left on the floor?
What the instruction sheet keeps
Automation works by capturing the explicit. A process is observed, broken into steps, and each step is specified, this much of that, heated to this point, held for this long. The specification is then handed to a machine that executes it identically every time. This is the great virtue of the method. It produces consistency, scale, and a result that does not depend on whether the operator slept well.
But the specification can only hold what was written down. And the writing-down is precisely the moment the tacit knowledge is lost. The baker’s two fingers in the dough cannot be transcribed. What gets transcribed instead is a proxy, a proving time, a hydration percentage, a target temperature, that approximates the judgement without containing it. On a good day the proxy is close enough. On an unusual day, when the flour is older or the room is humid, the baker would have adjusted and the machine does not, because the thing that would have noticed has been engineered out.
This is the quiet cost. Not that automated things are bad. That the variable the craftsperson was actually responding to, often the one that mattered most for quality, was never the variable that got specified. It couldn’t be. It was tacit. So it was discarded, and a measurable stand-in was kept in its place.
Trace, and the eye that reads it
Cold-process soapmaking turns on a moment called trace, when oils and lye have emulsified enough that the batter thickens and a drizzle drawn across the surface leaves a faint trail before sinking back. Trace is where scent and additives go in, where the texture of the final bar is partly decided. It can be reached too early or pushed too far.
You can instrument it. You can measure viscosity, hold temperatures, time the stir. And much of soapmaking is genuinely better for measurement, lye must be weighed to the gram, and no feel substitutes for that. But the reading of trace itself stays stubbornly in the eye and the wrist. The same recipe behaves differently with a different oil batch, a colder room, a fragrance that accelerates the reaction without warning. The maker sees it shifting and responds before a sensor would have flagged anything, because they are reading the whole behaviour at once rather than one isolated number.
That is the kind of knowing automation cannot inherit. Not because the technology is immature, but because the knowledge was never in a form that could be passed on except through doing. It is learned at the bench, in failures, in the difference between a batch that seized and one that didn’t.
Counting the real cost
None of this argues for inefficiency as a virtue. Slowness is not quality, and the hand is not holy. Plenty of handmade things are worse than their machined equivalents, made worse by exactly the inconsistency the machine removes.
The honest position is just to count properly. Automation gives consistency and scale and takes away the human variation, most of which was noise. But some of that variation was not noise. Some of it was a person responding, in real time, to conditions no specification anticipated, and that responsiveness was where a certain kind of quality lived.
When it goes, it tends to go silently. Nobody writes down what they could never write down. The loss does not appear on the ledger, because the thing lost was the part that was never legible in the first place.