Act 3 · The Machine · Station 06

The codifiability frontier

Watch tacit skills — driving, radiology, call-center expertise — migrate from human-only to machine-doable as AI milestones tick by.

Station 2 traced Hayek's shipper and estate agent — people whose knowledge was, in his account, tied so tightly to one person, one place, one fleeting moment that no report could ever capture it in time. Hayek treated that untransferability as a fact about the world. Brynjolfsson & Hitzig ask a sharper question in their §4.2: how much of what looked inalienable was really just uncodified — knowledge nobody had yet found a way to write down, extract, or train a model on? Three kinds of knowledge, and AI is moving the line on all three.

1. Explicit knowledge — already written down

Explicit knowledge is "knowledge that is easy to put into words or write down." AI does not need to extract it so much as digitize and absorb it at a cost and scale no human archivist could match: optical character recognition has driven the rapid digitization of enduring texts like the Vatican Apostolic Archive, handwritten in medieval Latin that many human readers struggled to understand, the reading of fast-passing license plate numbers for traffic management and enforcement, and the seamless uploading of physician notes into electronic health records.

Firmani et al., 2018; Lubna et al., 2021; Hsu et al., 2022 — B&H §4.2

2. Tacit knowledge — easy to act on, hard to explain

Explicit knowledge "stands in contrast to tacit knowledge (Polanyi, 1966), which is intuitively known – it's the knowledge that is easy to act on but hard to put into words." The paper's canonical examples are driving a car, language learning, and facial recognition — none of them things you could learn solely from reading a book. "AI systems have largely mastered the latter two forms of tacit knowledge in the last decade and are making progress on the first one."

A new marketplace now monetizes this directly: companies like Mercor connect specialists to AI firms to create bespoke training data — at the time of writing, their opportunities page calls for all manner of experts, from dermatologists at $270/hour to plant experts at $30–60/hour.

Mercor, 2025 — B&H §4.2

3. Machine-native knowledge — never a human's to begin with

"Beyond explicit knowledge and tacit knowledge, which have in the past been housed by human minds, machine learning has enabled the discovery or creation of entirely new forms of knowledge that no human mind – or set of human minds – could comprehend." Fraud detection, protein-folding predictions, preemptive firefighting allocation, high-frequency trading: none of these were ever tacit, because no human ever held them at all.

B&H §4.2

The frontier, mapped

Open a card to see the evidence. The first column is thin on purpose — the paper does not hand us a skill AI hasn't yet touched; watching cards leave that column is rather the whole story of §4.2.

Still becoming codified

  • Driving a car ＋

    By ingesting billions of miles of sensor data — camera feeds, LiDAR, radar, vehicle telemetry — AI now powers a fleet of 1,500 self-driving taxis (Waymos) completing over 250,000 rides a week in San Francisco, Austin, Phoenix and Los Angeles. The paper still frames this as AI "making progress" on driving, not having mastered it the way it has language or faces.

    Shepardson, 2025, in B&H §4.2

Recently codified

  • Language learning ＋

    Today's frontier large-language models are proficient in 50 languages — gaps in coverage reflect the uneven availability of training data rather than any fundamental barrier.

    OpenAI, 2024; Pava et al., 2025, in B&H §4.2

  • Facial recognition ＋

    State-of-the-art face-recognition systems, like Google's FaceNet, surpass 99.6 percent accuracy on the Labeled Faces in the Wild benchmark, outpacing human performance.

    Schroff et al., 2015, in B&H §4.2

  • Radiologists' visual expertise ＋

    Tools like MedGaze record thousands of eye-tracking sessions from practicing radiologists — logging where and how long experts fixate while reading X-rays — to train AI diagnostic assistants.

    Awasthi et al., 2025, in B&H §4.2

  • Call-center agent expertise ＋

    Training on millions of transcripts allowed a large language model to capture some of the best call center agents' expertise and make it available to newer and less skilled agents.

    Brynjolfsson et al., 2025, in B&H §4.2

Machine-native

  • Fraud detection ＋

    Pattern-recognition tools that flag fraudulent transactions no human reviewer could triage at that speed or scale.

    Dal Pozzolo et al., 2018, in B&H §4.2

  • Protein folding ＋

    Protein-folding predictions that suggest novel drug targets no human mind — or set of human minds — could comprehend.

    Jumper et al., 2021; Evans et al., 2022, in B&H §4.2

  • Preemptive firefighting ＋

    Anomaly-detection systems that allocate firefighting resources preemptively, ahead of the fire.

    Jain et al., 2020, in B&H §4.2

  • High-frequency trading ＋

    Algorithms that exploit microsecond price discrepancies — a timescale no human trader can act inside.

    Budish et al., 2015, in B&H §4.2

The name for this in the literature is Polanyi's paradox: we know more than we can tell, so for decades tacit skill sat outside the reach of automation almost by definition. The paper states plainly what changed:

"In each case, AI helps to overcome Polanyi's paradox (Autor, 2014) – we now no longer need to articulate a tacit knowledge or skill for it to be automated." Brynjolfsson & Hitzig, 2025, §4.2