Codified vs Tacit Knowledge Exposure

Sources#

Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence

Summary#

Every exposure instrument in Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated answers how much of an occupation AI reaches. This page holds a different question — which kind of knowledge it reaches — and the answer proposed for it: generative AI substitutes most effectively for knowledge that has been digitized, standardized and written down, while complementing the experiential knowledge that has not. The distinction (Acemoglu & Autor 2011, in this literature) is:

	Codified knowledge	Tacit knowledge
What it is	Formal, standardized, documented; explicit and transferable through text	Implicit, experience-based; hard to write down or teach in a classroom
How it is acquired	Education, textbooks, written procedures	Practice, mentorship, hands-on work, repeated exposure to real situations
Predicted AI relation	Substitute — it is exactly what a model trained on text has absorbed	Complement — the model supplies the codified half the practitioner already had to look up

The prediction that follows is a career-stage prediction rather than an occupational one: a 23-year-old's comparative advantage is recently-acquired codified knowledge, so AI arrives as a substitute for it; a 45-year-old's advantage is accumulated tacit knowledge, so the same tool arrives as a complement. Brynjolfsson, Chandar & Chen put it as AI "automating the checkable, process-intensive tasks that historically justified entry-level headcount, while increasing the leverage of experienced staff" (Ide 2025; Garicano & Rayo 2025).

This was the explanation attached to the age gradient for a year before anyone measured it. The August 2026 revision of "Canaries in the Coal Mine?" is the first measurement, and it is a good-faith one: the authors call it "a simple descriptive exercise," "non-causal and purely suggestive," and supply the footnote that undercuts half of it.

Evidence note. empirical — ADP payroll microdata, millions of workers, monthly through June 2026 (Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence). But the index on this page is not payroll data: it is an LLM's scoring of employer-written job descriptions, reported in an online appendix, with no standard errors, no causal design and one vintage. Treat the payroll series as measured and the codified/tacit axis laid over it as a constructed proxy with the failure modes recorded below.

How the index is actually built#

Appendix J, and it is worth reading before quoting the result:

Start from the 7,514 raw job titles in the ADP taxonomy, each with a job description ADP wrote.
Query GPT-5.4-mini at temperature 0, two separate API calls per job — one scoring codified knowledge 1–10, one scoring tacit knowledge 1–10. A shared system prompt teaches the model the distinction using O*NET knowledge domains, work activities and cross-functional skills as worked classifications, then the dimension-specific user prompt asks it to read the description, infer what knowledge each task draws on, and return JSON with a score and its reasoning.
Collapse to SOC level by unweighted average of the job titles mapping to each SOC code.
Code each SOC high or low on each dimension relative to the November 2022 employment-weighted mean of the SOC-level scores. That yields four quadrants: high-codified/high-tacit, high-codified/low-tacit, low-codified/high-tacit, low-codified/low-tacit.

Two things about this construction deserve to travel with any citation of it. First, the main text describes the proxies differently from the appendix — §2.5 says codified reliance is proxied "with its required level of formal education, supplemented by O*NET knowledge domains and work activities," and tacit reliance "with required work experience and on-the-job training, supplemented by" experiential domains, which reads as an O*NET-attribute index. Appendix J's operative procedure is an LLM reading prose job descriptions with O*NET categories supplied only as few-shot teaching material. Both descriptions are true of the intent; only the second is the instrument. Second, the unweighted title→SOC average and the employment-weighted high/low cut are both discretionary, and neither is varied in a robustness check.

What the payroll data shows under it#

Figure J.1 plots headcount by age band and quadrant, normalized to November 2022 = 1:

Early career (22–25) — the lines separate by codified, not by tacit. Both high-codified quadrants peak in 2024 and decline to roughly 0.95 by mid-2026; both low-codified quadrants stay at or just above 1.00. The tacit dimension barely orders anything in this age band.
Mid-career (41–49) and senior (50+) — the lines separate by tacit. At 41–49 the two high-tacit quadrants end near 1.11–1.14 against 1.06–1.08 for low-tacit, a spread of roughly 5–7 points of growth over the period. 35–40 shows the same ordering more weakly.
26–30 and 31–34 sit between, with no clean ordering on either dimension — the transition band.

The paper's own one-line summary: "For early-career workers, high-codified occupations grow more slowly than low-codified ones. For older workers, high-tacit occupations grow more quickly than low-tacit ones."

The asymmetry the authors flag, and why it is the whole point#

Footnote 21, which is easy to miss and which changes what this page is worth:

"The codified-knowledge gradient is not significant when controlling for college share quintile, demonstrating its overlap with education. The tacit-knowledge gradient for experienced workers, by contrast, survives this control."

So the two halves of the mechanism are not equally supported. The substitution half is collinear with schooling — "high codified" is close to "requires a degree," and the paper's main results already attenuate sharply under a college-share control (the most-exposed-quintile estimate for 22–25s moves from −0.18 with no controls to −0.09 with college share, Table 1 Panels A and C). The authors' own reading of that attenuation applies here too and cuts both ways: education may be a confounder, or it may be the channel, "since generative AI substitutes best for precisely the codified knowledge taught through formal schooling." An index built out of education cannot adjudicate between those two readings — which is exactly the fork this page exists to sharpen.

The complement half is the more robust finding, and it is the less-quoted one. Experienced workers in high-tacit occupations grow faster than experienced workers in low-tacit occupations, net of education. That is Returns to Expertise in Agentic Coding's thesis appearing in payroll data at population scale rather than in session telemetry.

The same age gradient, measured off usage instead of knowledge type#

The knowledge index is not the only instrument in the paper pointing this way, and the second one is independent of education by construction. Table 3 regresses occupation-level employment change (Nov 2022 → Jun 2026) jointly on standardized automation, complementarity and overall-usage exposure from the Anthropic Economic Index's automative/augmentative query split, separately by age band (March 2025 release; verified against pdftotext -layout):

Age	Automation	Complementarity	Overall usage
22–25	−0.098*** (0.018)	0.016 (0.021)	−0.029 (0.018)
26–30	−0.036*** (0.013)	0.010 (0.019)	0.000 (0.018)
31–34	−0.017 (0.011)	0.014 (0.015)	−0.017 (0.021)
35–40	−0.014 (0.011)	0.004 (0.014)	−0.023 (0.027)
41–49	−0.008 (0.010)	+0.024** (0.010)	−0.020 (0.025)
50+	−0.006 (0.009)	+0.015* (0.008)	−0.037** (0.016)

Read down the columns: the automation coefficient shrinks monotonically with age and the complementarity coefficient turns positive and significant only for the experienced. Percent change in employment per standard deviation of exposure. The pattern replicates on the pooled September 2025 / January 2026 / April 2026 releases (Panel B: −0.084*** at 22–25, +0.024*** at 41–49), where the complementarity coefficient for the young actually turns slightly negative.

This is the same shape as the knowledge index — substitution for the young, complementarity for the experienced — measured off what people send to a model rather than off what kind of knowledge a job needs. Two instruments agreeing is not two independent confirmations (both are occupation-level, both are laid over the same payroll panel), but the AEI measure owes nothing to years of schooling, which is the specific weakness of the knowledge index.

What this does and does not license#

It does not establish causation. Neither instrument does, and the authors say so twice.
It is an occupation-level statement, not a task-level one. Nothing here observes an individual's codified or tacit knowledge; "codified occupation" means an LLM judged the job description to lean that way. A senior and a junior in the same SOC code get the same score.
Tacit is defined negatively — as what cannot be written down — which makes it a moving target rather than a fixed reserve. That is Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated's experience gradient with the same soft spot: it rests on what has not yet been codified, not on what cannot be.
The complement finding is about the experienced, not about tacit work in general. High-tacit occupations do not grow faster for 22–25s; the interaction is with age, which is what makes it a statement about accumulated practice rather than about task content.

Connections#

The Tragedy of the Cognitive Commons — the mechanism this page measures is the premise under that page's argument: if AI substitutes for the codified knowledge that entry-level work used to pay for, the entry-level slot through which tacit knowledge was acquired disappears, and the commons stops regenerating. The relation is a dependency rather than an agreement — this page measures the substitution, and says nothing about whether the tacit knowledge of the next cohort ever forms
Returns to Expertise in Agentic Coding — the complement half at population scale and over four years, on a different instrument: where that page finds task-specific expertise amplifying an agent inside ~400K sessions, this finds experienced workers in tacit-heavy occupations growing faster than their low-tacit peers in payroll data. Both are the same claim about where human comparative advantage sits; neither shows it is durable
Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated — orthogonal axis. Every instrument there scores how much of an occupation AI reaches; this scores which kind of knowledge it reaches, which is why two occupations with identical exposure scores can have opposite age gradients. It also inherits that page's central problem in a new form — an LLM-scored index has an instrument's assumptions baked in exactly like a crowd-sourced rubric does
Task Saturation: Broad but Shallow AI Diffusion — the nearest sibling cut, and a different one: ATLAS measures depth (share of a reached occupation's tasks) and sorts tasks by the Autor–Thompson word-rarity expertise proxy, finding usage over-indexed on the lowest-expertise cognitive tasks. Expert/inexpert and codified/tacit are not the same axis — a task can be expert and fully codified (tax law) or inexpert and tacit (de-escalating an angry customer) — but they make the same prediction about who gets hit first, and this page supplies the labor-market outcome ATLAS's snapshot cannot
The Automation–Optimism Link — the same automative/augmentative split from the same index, read on sentiment instead of headcount: heavier delegators are more optimistic and report their skills growing more valuable, while the automation exposure of their occupation is what predicts employment decline for the young. Different unit (person vs occupation) and different outcome (felt vs paid), so no contradiction — but the pairing is the sharpest available statement of how far self-report and payroll can diverge on the same variable
Organizational Complements to AI — the complement side of the ledger at the firm level: what this page finds as a tacit-knowledge premium in occupational headcount is what that page treats as the organizational investment required to make AI augment rather than replace
Erik Brynjolfsson — the index's author, and the disambiguation page for the five different works of his the vault cites
Stanford Digital Economy Lab — the lab, its instruments and their standing caveats
AI Adoption in Scientific Work — the same mechanism argued for science without measurement: Jedlička's §7.1.2 says replacing doctoral and postdoctoral work removes the transmission of "tacit knowledge, methodological judgement, professional norms" that "cannot be fully formalised or codified" — this page's axis, stated by a philosopher of science who does not cite the labor evidence

Open Questions#

The codified gradient does not survive a college-share control and the tacit gradient does. Is codified/tacit a distinct axis at all, or an education index with a mechanism story attached? Falsifiable: rebuild the codified score from a substrate with no schooling content (O*NET on-the-job-training duration, apprenticeship and licensure requirements, work-experience zones) and re-run the same quadrant plot — if the early-career codified gradient survives, the axis is real; if it collapses, this page is measuring years of schooling.
The index is one LLM's reading of employer-written job descriptions at one vintage, with no inter-rater check of any kind. Does it reproduce under a different scorer, a different prompt, or human raters — and does the quadrant assignment move enough to flip the sign of any age band? Nothing in the source varies it.
Tacit knowledge is defined as what resists being written down, and the current wave of agentic tooling is an industrial effort to write work down — recorded workflows, skill files, agent memory. If capture succeeds, today's tacit premium is a codification backlog rather than a comparative advantage. Trigger to watch: the first occupation where a high-tacit score and a declining experienced-worker headcount appear together.

Sources#

Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence — Erik Brynjolfsson, Bharat Chandar & Ruyu Chen, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, Stanford Digital Economy Lab, revised August 2026 (140pp, empirical). Cited here for §2.5's "Mechanism Evidence: Codified versus Tacit Knowledge" paragraph and footnote 21 (the education-control asymmetry), Online Appendix Section J (index construction, prompts, SOC collapse, quadrant cut, Figure J.1), and Table 3 plus its surrounding prose (the automation/complementarity/usage regression by age). Parse note: Table 3's panel headers repeat across every cell — a row-span artifact that table-collapse flags 15 times across this document — but the coefficient cells are intact; every number quoted above was verified against pdftotext -layout (Table 3 and Table 1 reconcile cell-for-cell). Figure J.1 was read from the page image, not from text. Full parse ledger and evidence handling in Source Notes