The aggregate impact of AI on Norwegian employment

Exposure, the wage it travels with, and their interaction — to 2026Q1

Published

26 June 2026

The question

Exposure scores say what AI could touch. They do not say what happened to jobs. This page asks the aggregate question on Norwegian register-style data: once general-purpose AI arrived, did the most-exposed occupations — and especially the well-paid, exposed ones, where the economics says displacement should bite first — lose employment?

Data & method

Panel. Quarterly STYRK-08 occupational employment and monthly earnings, 2016Q1–2026Q1, from Statistics Norway (SSB tables 12542 and 11418); AI exposure from Eloundou et al. (2024).

The regression. One observation per occupation. The outcome is the Yagan (2019) percent change in employment relative to its pre-ChatGPT average, \[y_o = 100\,\frac{E_{o,\,2026Q1} - \bar E^{\text{pre}}_o}{\bar E^{\text{pre}}_o}, \qquad \bar E^{\text{pre}}_o = \text{mean employment over quarters } \le 2022Q3,\] a level-robust alternative to a log difference (it does not blow up when the base quarter is small or noisy). We build it up in three nested steps, each regressor standardised (so coefficients are per 1 SD, in percentage points of employment), with heteroskedasticity-robust SEs:

$y_o = \alpha + \beta_1\,\mathrm{Exp}_o + \varepsilon_o$ — the naive exposure regression.
add the pre-ChatGPT average log wage $\ln \bar w^{\text{pre}}_o$ (mean monthly wage over 2016Q1–2022Q3) as a control — pre-determined, so it is not contaminated by the post period.
add the interaction $\mathrm{Exp}_o \times \ln \bar w^{\text{pre}}_o$.

Why this order. Exposure and pay are correlated (high-exposure office work is well paid: corr ≈ 0.55). So a raw exposure coefficient is partly just wage. Putting wage in as a control lets OLS net that out; the interaction then asks the sharp comparative-advantage question — does exposure cost more jobs precisely where pay is high? (prediction $\beta_3<0$). Because all three are entered together, $\beta_3$ is the genuine, partialled interaction — there is no collinearity sleight of hand.

From one number to a path. The cross-section is a single long difference, so we also run the event-study (diff-in-diff) version: every coefficient gets a time subscript, estimated jointly with occupation and quarter fixed effects (ref. = 2022Q3, SE clustered by occupation), \[y_{o,t}=\alpha_o+\tau_t+\sum_{k}\big(\beta_{1,k}\,\mathrm{Exp}_o+\beta_{2,k}\ln\bar w_o +\beta_{3,k}\,\mathrm{Exp}_o\times\ln\bar w_o\big)\mathbf 1[t{=}k]+\varepsilon_{o,t}.\] We show it twice, exactly as in the deck: naive ($\beta_{1,k}$ alone — what people usually estimate) and joint (all three entered simultaneously, so $\beta_{1,k}$ is exposure net of wage and $\beta_{3,k}$ is the genuine heterogeneity-by-wage interaction).

Robustness (HonestDiD). A positive post-shock coefficient can be a real effect or the pre-existing trend continuing; Rambachan & Roth (2023) ask how big a post-period violation of parallel trends — as a fraction $\bar M$ of the largest pre-period wobble — is needed before zero re-enters the CI. That threshold, the breakdown $\bar M^\ast$, is the headline: small means fragile.

Results

Employment growth against exposure

Figure 1: Each dot is a STYRK-08 occupation; **area is employment in 2026Q1** (with a size floor so small occupations stay visible). The most-exposed occupations are ringed and labelled. If exposure destroyed jobs the cloud would slope down — it does not.

The cloud is flat-to-slightly-positive. Note two things, because they answer the obvious question “where are the 90–100%-exposure software jobs?”: the most-exposed occupations (programmers, database/data-entry, payroll) sit at exposure 0.86–0.97 (ringed) but employ only a few hundred to a few thousand each, so they barely move aggregate employment; and the single 100%-exposure occupation (“Kodere mv.”) is dropped, because SSB suppresses its wage (≈ 6–14 employed) so it has no wage to control for.

The regression: exposure, then control for wage, then the interaction

Effect on employment (percentage points), per 1 SD; robust SE in parentheses; *** p<.01, ** p<.05, * p<.1.

Employment, % vs pre-ChatGPT avg (Yagan)	(1) exposure only	(2) + log pre-wage	(3) + interaction
AI exposure (β₁)	+4.4 (1.6)***	+0.9 (1.9)	+1.5 (1.8)
log pre-ChatGPT wage (β₂)		+6.4 (1.8)***	+5.9 (1.8)***
exposure × log wage (β₃)			+2.8 (2.0)
R²	0.025	0.063	0.068
N (occupations)	357	357	357

Read it left to right — this is the whole story:

(1) Naive. On its own, AI exposure looks good for jobs: +4.4 (1.6)***% per SD, significant. Exposed occupations grew. (The same lesson as the deck’s event-study exposure and comparative-advantage slides — both ≈ +3 pp.)
(2) Control for pre-ChatGPT wage. The exposure coefficient collapses to +0.9 (1.9) — indistinguishable from zero — while the wage control takes a large, significant +6.4 (1.8)*. The naive “exposure effect” was wage in disguise**: high-exposure occupations are well paid, and well-paid occupations grew. OLS nets it out.
(3) Add the interaction. Exposure stays at +1.5 (1.8) and the genuine interaction is +2.8 (2.0) — not significant. There is no extra job loss in high-wage exposed occupations: the sharp comparative-advantage prediction ($\beta_3<0$) does not show up.

The number that matters: total exposure effect by wage level

Because the effect of exposure now depends on wage, the quantity to report is the total exposure effect for an occupation, $\beta_1+\beta_3\cdot(\text{wage})$:

Total effect of +1 SD exposure	Employment
in a high-wage occupation (β₁ + β₃)	+4.3% [-1.7, +10.3]
in a low-wage occupation (β₁ − β₃)	-1.3% [-6.0, +3.4]

Even at the top of the pay distribution — where displacement was predicted to bite hardest — the total employment effect of AI exposure is +4.3% [-1.7, +10.3], indistinguishable from zero. No displacement, including where the theory said to look for it.

The naive event study: exposure only

Figure 2: Event study of the naive exposure effect (β₁, per 1 SD): employment (% vs the occupation’s pre-ChatGPT average, Yagan 2019) on exposure × quarter, occupation + quarter fixed effects, 95% CIs clustered by occupation. Dashed amber line = ChatGPT (2022Q4); green band = average post-shock effect; dashed pre-trend slope fitted on the pre-period.

This is what people usually estimate — and on its own it looks like good news: +3.11 pp per SD post-ChatGPT. But the path shows a clear upward pre-trend (+0.107 pp/quarter) — exposed occupations were already drifting up before ChatGPT — which is exactly why the naive number cannot be read causally.

The same event study with heterogeneity by wage: the joint decomposition

The deck’s causal centerpiece. Exposure, the log wage, and their (centered) interaction each get a full set of quarter interactions, entered simultaneously — so the exposure path is net of wage, and the interaction path is the genuine does-exposure-bite-where-pay-is-high object ($\beta_3 < 0$ is the displacement prediction).

Figure 3: Joint event study: exposure β₁,ₜ **net of wage and the interaction** (per 1 SD). Read it as a difference-in-differences: how employment in a more-exposed occupation moves around ChatGPT relative to a less-exposed one, net of quarter shocks and fixed occupation differences.

Figure 4: Joint event study: the interaction β₃,ₜ = Exp × ln w̄ (per 1 SD) — the comparative-advantage / heterogeneity-by-wage path. The sharp displacement prediction is β₃ < 0.

Net of the wage it travels with, exposure does nothing: the average post estimate is +0.3% [-1.5, +2.0]. And the interaction — the displacement object — is -1.6% [-4.0, +0.9], flat through the shock. The prediction $\beta_3<0$ does not appear: even the well-paid, exposed occupations show no extra decline. (The comparative-advantage signal in a one-regressor fit is carried by the wage main effect, +4.8% [+1.7, +7.9].)

Causal interpretation: HonestDiD on naive vs joint

Per-SD average post effects with Rambachan–Roth (2023) relative-magnitudes robustness — the same table as the deck:

	avg post (pp/SD)	95% CI	HonestDiD 95% CI (M̄=1)	breakdown M̄*
Naive exposure only (β₁)	+3.11	[+1.22, +4.99]	[-26.4, +32.6]	0.04
Joint fit: exposure (β₁)	+0.27	[-1.51, +2.04]	[-8.4, +9.0]	n.s.
Joint fit: interaction (β₃)	-1.57	[-4.04, +0.90]	[-32.4, +29.3]	n.s.

The only positive-looking estimate is the naive exposure effect (+3.1 pp/SD). Its breakdown is $\bar M^\ast \approx 0.04$: a post-shock deviation from parallel trends just 4% as large as the worst pre-ChatGPT wobble already puts zero back in the interval — not robust.
In the joint fit, exposure (β₁) and the genuine interaction (β₃) are already ≈ 0 — there is no positive effect left for a pre-trend to overturn.
Causal verdict: no robust effect in either direction — no displacement, no dividend.

Bottom line

The naive exposure effect is wage in disguise. +4.4 (1.6)**% per SD on its own → +0.9 (1.9)% once pre-ChatGPT wage is controlled; in the joint event study, exposure net of wage is +0.3% [-1.5, +2.0]. What grew were well-paid* occupations, exposed or not.
No displacement, including where predicted. The heterogeneity-by-wage interaction is -1.6% [-4.0, +0.9] in the joint event study (+2.8 (2.0)% in the cross-section, n.s. both ways); the total exposure effect for a high-wage occupation is +4.3% [-1.7, +10.3] — zero. Aggregate employment in these 357 occupations rose 0.3% from 2022Q3 to 2026Q1 (2,767,939 employed in 2026Q1).
And the naive positive effect is not robust either. A strong pre-trend drives it: HonestDiD breakdown $\bar M^\ast \approx 0.04$. Causal verdict: no robust effect in either direction. Displacement, if it comes, should surface first in hiring and entry-level flows, which a stock panel understates — the next test.

Take this study into the AI Impact Analyst →

The analyst opens with every exhibit of this study loaded in the Exhibit picker, and can edit each one on request — change the specification, the standard errors, the wage definition, or the variables — and the revised figure or table appears directly in the chat.

References

Lindenlaub, I., Oh, R., Rodríguez, M. A. & Veldkamp, L. (2026). Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage.
Eloundou, T., Manning, S., Mishkin, P. & Rock, D. (2024). GPTs are GPTs: Labor Market Impact Potential of LLMs.
Rambachan, A. & Roth, J. (2023). A More Credible Approach to Parallel Trends. Review of Economic Studies.
Statistics Norway (SSB), StatBank via PxWebApi v2 — tables 12542 (employment) and 11418 (monthly earnings).

--- title: "The aggregate impact of AI on Norwegian employment" subtitle: "Exposure, the wage it travels with, and their interaction — to 2026Q1" date: today date-format: "D MMMM YYYY" toc: true engine: jupyter format: html: code-tools: true code-links: false grid: body-width: 1040px --- ```{python} #| label: setup #| include: false import numpy as np import pandas as pd import matplotlib.pyplot as plt import koja_impact as ki KOJA = ki.KOJA plt.rcParams.update({ "figure.facecolor": "none", "axes.facecolor": "none", "savefig.facecolor": "none", "savefig.transparent": True, "text.color": KOJA["mist"], "axes.labelcolor": KOJA["mist"], "axes.edgecolor": KOJA["edge"], "xtick.color": KOJA["muted"], "ytick.color": KOJA["muted"], "axes.titlecolor": KOJA["heading"], "axes.grid": True, "grid.color": KOJA["muted"], "grid.alpha": 0.15, "font.size": 11, }) cs, panel, qs, qidx, ref_i, shock_i = ki.load() R = ki.nested_cross_section() # the headline build-up regression es_exp = ki.event_study(panel, qs, "T_exp") # dynamics for the naive exposure effect he = ki.honest_rm(es_exp["path"], es_exp["theta"], es_exp["se"], es_exp["J"], qs) jt = ki.joint_decomposition(panel, qs) # the deck's joint event-study decomposition he_j = {tr: ki.honest_rm(jt[tr]["path"], jt[tr]["theta"], jt[tr]["se"], jt[tr]["J"], qs) for tr in ("T_exp", "T_xw")} def md(s): from IPython.display import Markdown; return Markdown(s) def stars(p): return "***" if p < .01 else "**" if p < .05 else "*" if p < .1 else "" def cell(sp, k): # "+4.4 (1.6)***" or blank if k not in sp: return "" # coefficients already in pp (Yagan outcome) b, se, p = sp[k] return f"{b:+.1f} ({se:.1f}){stars(p)}" def ci(d): return f"{d['theta']:+.1f}% [{d['theta']-1.96*d['se']:+.1f}, {d['theta']+1.96*d['se']:+.1f}]" def hrow(lab, r, hd): # row of the causal-interpretation table sig = abs(r["theta"]) > 1.96 * r["se"] bd = f"{hd['breakdown']:.2f}" if sig else "n.s." return (f"| {lab} | {r['theta']:+.2f} | [{r['theta']-1.96*r['se']:+.2f}, " f"{r['theta']+1.96*r['se']:+.2f}] | [{hd['lo']:+.1f}, {hd['hi']:+.1f}] | {bd} |") T = R["table"]; C1, C2, C3 = "(1) exposure only", "(2) + log pre-wage", "(3) + interaction" N = R["N"] agg = panel.groupby("quarter").emp.sum() agg_growth = (agg["2026K1"] / agg["2022K3"] - 1) * 100 emp_2026 = agg["2026K1"] ``` ## The question Exposure scores say what AI *could* touch. They do not say what happened to **jobs**. This page asks the aggregate question on Norwegian register-style data: once general-purpose AI arrived, did the most-exposed occupations — and especially the **well-paid, exposed** ones, where the economics says displacement should bite first — *lose* employment? ## Data & method **Panel.** Quarterly STYRK-08 occupational employment and monthly earnings, **2016Q1–2026Q1**, from Statistics Norway (SSB tables [12542](https://www.ssb.no/en/statbank/table/12542) and [11418](https://www.ssb.no/en/statbank/table/11418)); AI exposure from **Eloundou et al. (2024)**. **The regression.** One observation per occupation. The outcome is the **Yagan (2019) percent change** in employment relative to its pre-ChatGPT average, $$y_o = 100\,\frac{E_{o,\,2026Q1} - \bar E^{\text{pre}}_o}{\bar E^{\text{pre}}_o}, \qquad \bar E^{\text{pre}}_o = \text{mean employment over quarters } \le 2022Q3,$$ a level-robust alternative to a log difference (it does not blow up when the base quarter is small or noisy). We build it up in three nested steps, each regressor **standardised** (so coefficients are *per 1 SD*, in percentage points of employment), with heteroskedasticity-robust SEs: 1. $y_o = \alpha + \beta_1\,\mathrm{Exp}_o + \varepsilon_o$ — the naive exposure regression. 2. add the **pre-ChatGPT average log wage** $\ln \bar w^{\text{pre}}_o$ (mean monthly wage over 2016Q1–2022Q3) as a control — *pre-determined*, so it is not contaminated by the post period. 3. add the **interaction** $\mathrm{Exp}_o \times \ln \bar w^{\text{pre}}_o$. **Why this order.** Exposure and pay are correlated (high-exposure office work is well paid: corr ≈ `{python} f"{R['corr_exp_wage']:.2f}"`). So a raw exposure coefficient is partly just *wage*. Putting wage in as a control lets OLS **net that out**; the interaction then asks the sharp comparative-advantage question — *does exposure cost more jobs precisely where pay is high?* (prediction $\beta_3<0$). Because all three are entered together, $\beta_3$ is the genuine, partialled interaction — there is no collinearity sleight of hand. **From one number to a path.** The cross-section is a single long difference, so we also run the **event-study (diff-in-diff) version**: every coefficient gets a time subscript, estimated jointly with occupation and quarter fixed effects (ref. = 2022Q3, SE clustered by occupation), $$y_{o,t}=\alpha_o+\tau_t+\sum_{k}\big(\beta_{1,k}\,\mathrm{Exp}_o+\beta_{2,k}\ln\bar w_o +\beta_{3,k}\,\mathrm{Exp}_o\times\ln\bar w_o\big)\mathbf 1[t{=}k]+\varepsilon_{o,t}.$$ We show it twice, exactly as in the deck: **naive** ($\beta_{1,k}$ alone — what people usually estimate) and **joint** (all three entered simultaneously, so $\beta_{1,k}$ is exposure *net of wage* and $\beta_{3,k}$ is the genuine heterogeneity-by-wage interaction). **Robustness (HonestDiD).** A positive post-shock coefficient can be a real effect *or* the pre-existing trend continuing; Rambachan & Roth (2023) ask how big a post-period violation of parallel trends — as a fraction $\bar M$ of the largest pre-period wobble — is needed before zero re-enters the CI. That threshold, the **breakdown $\bar M^\ast$**, is the headline: small means fragile. ## Results ### Employment growth against exposure ```{python} #| label: fig-growth-exposure #| echo: false #| fig-cap: "Each dot is a STYRK-08 occupation; **area is employment in 2026Q1** (with a size floor so small occupations stay visible). The most-exposed occupations are ringed and labelled. If exposure destroyed jobs the cloud would slope down — it does not." fig, ax = plt.subplots(figsize=(9.0, 5.0)) slope, r2 = ki.draw_scatter(ax, cs) plt.tight_layout(); plt.show() ``` ```{python} #| output: asis #| echo: false md(f"""The cloud is flat-to-slightly-positive. Note two things, because they answer the obvious question *"where are the 90–100%-exposure software jobs?"*: the most-exposed occupations (programmers, database/data-entry, payroll) sit at exposure 0.86–0.97 (ringed) but **employ only a few hundred to a few thousand each**, so they barely move aggregate employment; and the single 100%-exposure occupation ("Kodere mv.") is **dropped**, because SSB suppresses its wage (≈ 6–14 employed) so it has no wage to control for.""") ``` ### The regression: exposure, then control for wage, then the interaction ```{python} #| output: asis #| echo: false md(f"""Effect on employment (percentage points), per 1 SD; robust SE in parentheses; *** p<.01, ** p<.05, * p<.1. | Employment, % vs pre-ChatGPT avg (Yagan) | {C1} | {C2} | {C3} | |---|:--:|:--:|:--:| | **AI exposure** (β₁) | {cell(T[C1],'exposure')} | {cell(T[C2],'exposure')} | {cell(T[C3],'exposure')} | | **log pre-ChatGPT wage** (β₂) | | {cell(T[C2],'logwage')} | {cell(T[C3],'logwage')} | | **exposure × log wage** (β₃) | | | {cell(T[C3],'interaction')} | | R² | {T[C1]['R2']:.3f} | {T[C2]['R2']:.3f} | {T[C3]['R2']:.3f} | | N (occupations) | {N} | {N} | {N} | Read it left to right — this is the whole story: - **(1) Naive.** On its own, AI exposure looks *good* for jobs: **{cell(T[C1],'exposure')}**% per SD, significant. Exposed occupations grew. (The same lesson as the deck's event-study exposure and comparative-advantage slides — both ≈ +3 pp.) - **(2) Control for pre-ChatGPT wage.** The exposure coefficient **collapses to {cell(T[C2],'exposure')}** — indistinguishable from zero — while the wage control takes a large, significant **{cell(T[C2],'logwage')}**. The naive "exposure effect" was **wage in disguise**: high-exposure occupations are well paid, and well-paid occupations grew. OLS nets it out. - **(3) Add the interaction.** Exposure stays at **{cell(T[C3],'exposure')}** and the genuine interaction is **{cell(T[C3],'interaction')}** — not significant. There is **no extra job loss in high-wage exposed occupations**: the sharp comparative-advantage prediction ($\\beta_3<0$) does not show up.""") ``` ### The number that matters: total exposure effect by wage level ```{python} #| output: asis #| echo: false md(f"""Because the effect of exposure now depends on wage, the quantity to report is the **total** exposure effect for an occupation, $\\beta_1+\\beta_3\\cdot(\\text{{wage}})$: | Total effect of +1 SD exposure | Employment | |---|:--:| | in a **high-wage** occupation (β₁ + β₃) | **{ci(R['high_wage'])}** | | in a low-wage occupation (β₁ − β₃) | {ci(R['low_wage'])} | Even at the top of the pay distribution — where displacement was predicted to bite hardest — the total employment effect of AI exposure is **{ci(R['high_wage'])}**, indistinguishable from zero. No displacement, including where the theory said to look for it.""") ``` ### The naive event study: exposure only ```{python} #| label: fig-event-exposure #| echo: false #| fig-cap: "Event study of the naive exposure effect (β₁, per 1 SD): employment (% vs the occupation's pre-ChatGPT average, Yagan 2019) on exposure × quarter, occupation + quarter fixed effects, 95% CIs clustered by occupation. Dashed amber line = ChatGPT (2022Q4); green band = average post-shock effect; dashed pre-trend slope fitted on the pre-period." fig, ax = plt.subplots(figsize=(10.5, 4.6)) ki.draw_eventstudy(ax, es_exp, qs, ref_i, shock_i, f"Naive exposure effect over time (β₁; N={len(cs)})", KOJA["purple"]) plt.tight_layout(); plt.show() ``` ```{python} #| output: asis #| echo: false md(f"""This is *what people usually estimate* — and on its own it looks like good news: **{es_exp['theta']:+.2f} pp per SD** post-ChatGPT. But the path shows a clear upward **pre-trend** ({es_exp['pre']:+.3f} pp/quarter) — exposed occupations were already drifting up before ChatGPT — which is exactly why the naive number cannot be read causally.""") ``` ### The same event study with heterogeneity by wage: the joint decomposition The deck's causal centerpiece. Exposure, the log wage, and their (centered) interaction each get a **full set of quarter interactions, entered simultaneously** — so the exposure path is *net of wage*, and the interaction path is the genuine *does-exposure-bite-where-pay-is-high* object ($\beta_3 < 0$ is the displacement prediction). ```{python} #| label: fig-event-joint-exposure #| echo: false #| fig-cap: "Joint event study: exposure β₁,ₜ **net of wage and the interaction** (per 1 SD). Read it as a difference-in-differences: how employment in a more-exposed occupation moves around ChatGPT relative to a less-exposed one, net of quarter shocks and fixed occupation differences." fig, ax = plt.subplots(figsize=(10.5, 4.6)) ki.draw_eventstudy(ax, jt["T_exp"], qs, ref_i, shock_i, f"Joint: exposure β₁,ₜ net of wage & interaction (N={len(cs)})", KOJA["purple"]) plt.tight_layout(); plt.show() ``` ```{python} #| label: fig-event-joint-interaction #| echo: false #| fig-cap: "Joint event study: the interaction β₃,ₜ = Exp × ln w̄ (per 1 SD) — the comparative-advantage / heterogeneity-by-wage path. The sharp displacement prediction is β₃ < 0." fig, ax = plt.subplots(figsize=(10.5, 4.6)) ki.draw_eventstudy(ax, jt["T_xw"], qs, ref_i, shock_i, f"Joint: interaction β₃,ₜ = Exp × ln w̄, net of main effects (N={len(cs)})", KOJA["fjord"]) plt.tight_layout(); plt.show() ``` ```{python} #| output: asis #| echo: false md(f"""Net of the wage it travels with, exposure does **nothing**: the average post estimate is **{ci(jt['T_exp'])}**. And the interaction — the displacement object — is **{ci(jt['T_xw'])}**, flat through the shock. The prediction $\\beta_3<0$ **does not appear**: even the well-paid, exposed occupations show no extra decline. (The comparative-advantage signal in a one-regressor fit is carried by the **wage main effect**, {ci(jt['T_wage'])}.)""") ``` ### Causal interpretation: HonestDiD on naive vs joint ```{python} #| output: asis #| echo: false md("\n".join([ "Per-SD average post effects with Rambachan–Roth (2023) relative-magnitudes robustness — the same table as the deck:", "", "| | avg post (pp/SD) | 95% CI | HonestDiD 95% CI (M̄=1) | breakdown M̄* |", "|---|:--:|:--:|:--:|:--:|", hrow("*Naive* exposure only (β₁)", es_exp, he), hrow("Joint fit: exposure (β₁)", jt["T_exp"], he_j["T_exp"]), hrow("Joint fit: interaction (β₃)", jt["T_xw"], he_j["T_xw"]), "", f"""- The only positive-looking estimate is the **naive** exposure effect ({es_exp['theta']:+.1f} pp/SD). Its breakdown is **$\\bar M^\\ast \\approx {he['breakdown']:.2f}$**: a post-shock deviation from parallel trends just **{he['breakdown']*100:.0f}% as large as the worst pre-ChatGPT wobble** already puts zero back in the interval — **not robust**. - In the **joint** fit, exposure (β₁) and the genuine interaction (β₃) are **already ≈ 0** — there is no positive effect left for a pre-trend to overturn. - **Causal verdict: no robust effect in either direction** — no displacement, no dividend.""", ])) ``` ## Bottom line ```{python} #| output: asis #| echo: false md(f"""- **The naive exposure effect is wage in disguise.** {cell(T[C1],'exposure')}% per SD on its own → {cell(T[C2],'exposure')}% once pre-ChatGPT wage is controlled; in the joint event study, exposure net of wage is {ci(jt['T_exp'])}. What grew were *well-paid* occupations, exposed or not. - **No displacement, including where predicted.** The heterogeneity-by-wage interaction is {ci(jt['T_xw'])} in the joint event study ({cell(T[C3],'interaction')}% in the cross-section, n.s. both ways); the total exposure effect for a high-wage occupation is {ci(R['high_wage'])} — zero. Aggregate employment in these {N} occupations **{'rose' if agg_growth>0 else 'fell'} {abs(agg_growth):.1f}%** from 2022Q3 to 2026Q1 ({emp_2026:,.0f} employed in 2026Q1). - **And the naive positive effect is not robust either.** A strong pre-trend drives it: HonestDiD breakdown $\\bar M^\\ast \\approx {he['breakdown']:.2f}$. **Causal verdict: no robust effect in either direction.** Displacement, if it comes, should surface first in **hiring and entry-level flows**, which a stock panel understates — the next test.""") ``` ```{=html} <div style="border:1px solid #3a4860;border-radius:10px;background:rgba(111,174,142,.06);padding:1rem 1.1rem;margin:0.2rem 0 1.6rem 0"> <a href="analysis.html?expert=ai-impact&study=aggregate-impact" style="display:inline-block;background:#cf9a5c;color:#070b12;font-weight:600;padding:.55rem 1.1rem;border-radius:8px;text-decoration:none"> Take this study into the AI Impact Analyst →</a> <p style="color:#8292ad;font-size:.86rem;margin:.5rem 0 0 0"> The analyst opens with <b>every exhibit of this study loaded</b> in the Exhibit picker, and can edit each one on request — change the specification, the standard errors, the wage definition, or the variables — and the revised figure or table appears directly in the chat.</p> </div> ``` ## References - Lindenlaub, I., Oh, R., Rodríguez, M. A. & Veldkamp, L. (2026). *Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage.* - Eloundou, T., Manning, S., Mishkin, P. & Rock, D. (2024). *GPTs are GPTs: Labor Market Impact Potential of LLMs.* - Rambachan, A. & Roth, J. (2023). *A More Credible Approach to Parallel Trends.* Review of Economic Studies. - Statistics Norway (SSB), StatBank via PxWebApi v2 — tables 12542 (employment) and 11418 (monthly earnings).