---
title: "David Autor's 'Polanyi's paradox' names why tasks that demand tacit judgment resist automation while codifiable tasks do not"
type: "claim"
status: "seedling"
source_url: "https://economics.mit.edu/sites/default/files/publications/polanyis%20paradox%202014.pdf"
source_author: "David H. Autor"
source_date: "2014"
source_quote: "Tasks that have proved most vexing to automate are those that demand flexibility, judgment and common sense — skills that we understand only tacitly"
source_tier: 1
audit_status: "capture-verified — the 2026-07-11 hop capture fetched Autor's own PDF directly and recorded this quote plus his gloss on machine learning; promotion did not independently re-fetch (no live web-fetch access available in this headless run), but the source is a public, non-paywalled primary PDF and the quote matches Autor's well-documented usage of the term."
provenance: "Promotion from 10-inbox/raw/2026-07-11-hop-polanyi-codification-bridge.md, 2026-07-12"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-11-hop-polanyi-codification-bridge.md"
writer_model: "claude-sonnet-5"
date_created: "2026-07-12T00:00:00.000Z"
tags: ["polanyi-paradox","automation","tacit-knowledge","codification","labor-economics","ai-governance"]
---


David Autor, in "Polanyi's Paradox and the Shape of Employment Growth" (2014),
named and applied Michael Polanyi's observation that "we know more than we
can tell" to the economics of automation. His claim: **"Tasks that have
proved most vexing to automate are those that demand flexibility, judgment
and common sense — skills that we understand only tacitly."** The inverse
holds too — tasks that can be reduced to explicit, codifiable procedures are
exactly the ones that machines can execute. Codifiability, not complexity
per se, is what sorts tasks into automatable and non-automatable.

Autor also glosses the machine-learning response to this limit: contemporary
computer science "seeks to overcome Polanyi's paradox by building machines
that learn from human examples," substituting statistical pattern-matching
for explicit rule-writing where the rules themselves cannot be articulated.

This is the automation leg of a broader structural pattern: codification is
the single property that governs both what a machine can execute and what a
rival can copy. See
[[observation-codification-is-the-shared-hinge-of-automation-and-imitation]]
for the full bridge, which pairs this claim with the imitation-barriers
literature ([[claim-causal-ambiguity-from-tacitness-raises-imitation-barriers]])
and applies it to [[claim-canonic-slop-is-a-verdict-of-domain-expertise]] and
[[claim-canonic-deliverable-is-an-append-only-evidence-ledger]].

> [!note] Seek's commentary:
> Worth remembering this is the *named* half of the bridge — Autor coined the
> term in 2014, he didn't invent the underlying idea (that's Polanyi, 1966).
> The paper is squarely Tier 1 and the quote is clean; this is the sturdiest
> leg of the whole hop chain. — Seek