Polanyi's Paradox

Definition

Philosopher Michael Polanyi's observation that "we know more than we can tell" describes skills humans perform fluently but cannot articulate as explicit rules — the pattern recognition involved in recognizing a face, grasping context in language, or driving through traffic. Because early approaches to automation required a human expert to codify a task into explicit if-then rules (an algorithm), any task resting on this kind of tacit knowledge was, for decades, presumed permanently safe from computers — not because it was intrinsically hard, but because no one could write down how they did it.

In the Book

The book presents economists Frank Levy and Richard Murnane's 2004 framework distinguishing rule-based tasks (like mortgage approval, reducible to explicit criteria) from pattern-recognition tasks that "cannot be boiled down to rules," citing their example of a truck driver navigating a left turn through oncoming traffic, signals, and pedestrians — knowledge the authors argued was "enormously difficult" to embed in software. Brynjolfsson and McAfee recount being persuaded by this argument in 2004, the same year DARPA's first Grand Challenge for autonomous vehicles ended in the "Debacle in the Desert" (the best car covered under 5 percent of a 150-mile course before crashing). But by 2010 Google's self-driving cars were logging highway miles, because the successful approach abandoned trying to codify driving as explicit rules and instead had machines learn statistical patterns directly from vast quantities of driving data — sidestepping the paradox rather than solving it by articulation. The book links this to Moravec's paradox (sensorimotor and pattern-recognition skills that are easy for humans prove hardest for machines, while abstract calculation that is hard for humans is easy for machines).

Why It Matters

The paradox reframes "can this be automated?" from a question about task difficulty to a question about whether the relevant knowledge is explicit-and-articulable or tacit-and-inarticulable — and it flags that the tacit/rule-based boundary is not fixed. Approaches that learn from examples rather than requiring an expert to state rules (as happened with driving, and later with image recognition and translation) can cross the boundary that pure rule-coding could not, which is a general diagnostic for predicting which currently "safe" human skills are actually vulnerable to a change in method rather than a change in task.