How do engineers overcome the limits of tacit knowledge in automation?

Autor names "Polanyi's paradox" after economist, philosopher and chemist Michael Polanyi's 1966 observation that "we know more than we can tell" — the gap between what people can do (identify a bird from a glimpse, write a persuasive paragraph, break an egg over a bowl) and what they can state as explicit rules. The paradox, Autor argues, "helps to explain what has not yet been accomplished" in automation and "illuminates the paths by which more will ultimately be accomplished." He names two such paths: "environmental control," which "circumvents Polanyi's paradox by regularizing the environment, so that comparatively inflexible machines can function semi-autonomously," and machine learning, which "inverts Polanyi's paradox" by having machines "infer tacit rules from context, abundant data, and applied statistics" rather than being taught rules nobody can state.

For environmental control, Autor contrasts assembly-line robots that install windshields on new cars with human technicians who still do aftermarket windshield replacement, because removing and refitting a broken windshield "demand[s] more real-time adaptability than any contemporary robot can cost-effectively approach" outside a simplified factory environment. For machine learning, he walks through why no rules-based program can define "chair" without misclassifying tables or excluding backless chairs, while "any grade-school child could perform this task with high accuracy" — a gap machine learning closes by training on labeled examples rather than specified features. He tempers both paths with failure cases: IBM Watson's Jeopardy win included a confidently wrong answer ("Toronto"), and a 2012 neural network needed "16,000 processors to identify images of cats on YouTube." Machine learning, Autor cautions, "may have fundamental problems with reasoning about 'purposiveness' and intended uses" — invoking Sagan's line that making an apple pie from scratch requires first inventing the universe.

None of the nearest notes in this library name Polanyi's paradox or this two-path taxonomy directly; the closer fit is the companion note Does automating some tasks make other human tasks more valuable?, drawn from the same essay, which explains why automated and human tasks end up complementing each other once some tasks are automated. This note answers a prior question the other assumes: which tasks get automated at all. It also qualifies How does automating junior work destroy verification oversight? and What becomes scarce when AI execution grows cheap?: both describe verification becoming scarce as automation accelerates, but Autor's machine-learning path is itself a verification problem in miniature — a system trained on labeled "ground truth" still cannot "tell" a programmer why it classified something as it did, the same opacity those notes locate in agent oversight, stated here about object recognition rather than AI agents.

Autor's examples — ATMs, warehouse pickers, Watson, an image classifier identifying cats — predate the generative and agentic AI systems this library mostly concerns; the essay does not claim its taxonomy covers large language models, and its 2015 confidence that reasoning about "purposiveness" would remain hard for machine learning predates evidence either way about LLM agents. The excerpt also does not quantify how much of the task space environmental control versus machine learning will eventually cover — it offers the two paths as the available engineering strategies of 2015, not a forecast of which will win or how far either will reach.